improving the reliability of commodity operating...
TRANSCRIPT
Improving the Reliability of Commodity Operating Systems
Michael M. Swift
A dissertation submitted in partial fulfillmentof the requirements for the degree of
Doctor of Philosophy
University of Washington
2005
Program Authorized to Offer Degree: Computer Science and Engineering
University of WashingtonGraduate School
This is to certify that I have examined this copy of a doctoral dissertation by
Michael M. Swift
and have found that it is complete and satisfactory in all respects,and that any and all revisions required by the final
examining committee have been made.
Co-Chairs of Supervisory Committee:
Henry M. Levy
Brian N. Bershad
Reading Committee:
Henry M. Levy
Brian N. Bershad
John Zahorjan
Date:
In presenting this dissertation in partial fulfillment of the requirements for the doctoral degree at
the University of Washington, I agree that the Library shall make its copies freely available for
inspection. I further agree that extensive copying of this dissertation is allowable only for scholarly
purposes, consistent with “fair use” as prescribed in the U.S. Copyright Law. Requests for copying
or reproduction of this dissertation may be referred to Proquest Information and Learning, 300 North
Zeeb Road, Ann Arbor, MI 48106-1346, to whom the author has granted “the right to reproduce and
sell (a) copies of the manuscript in microform and/or (b) printed copies of the manuscript made from
microform.”
Signature
Date
University of Washington
Abstract
Improving the Reliability of Commodity Operating Systems
Michael M. Swift
Co-Chairs of Supervisory Committee:
Professor Henry M. LevyComputer Science and Engineering
Associate Professor Brian N. BershadComputer Science and Engineering
This dissertation presents an architecture and mechanism for improving the reliability of com-
modity operating systems by enabling them to tolerate component failures.
Improving reliability is one of the greatest challenges for commodity operating systems, such as
Windows and Linux. System failures are commonplace in all domains: in the home, in the server
room, and in embedded systems, where the existence of the OS itself is invisible. At the low end,
failures lead to user frustration and lost sales. At the high end, an hour of downtime from a system
failure can result in losses in the millions.
Device drivers are a leading cause of failure in commodity operating systems. In Windows
XP, for example, device drivers cause 85% of reported failures [160]. In Linux, the frequency of
coding errors is up to seven times higher for device drivers than for the rest of the kernel [45].
Despite decades of research in reliability and extensible operating system technology, little of this
technology has made it into commodity systems. One major reason for the lack of adoption is
that these research systems require that the operating system, drivers, applications, or all three be
rewritten to take advantage of the technology.
In this dissertation I present a layered architecture for tolerating the failure of existing drivers
within existing operating system kernels. My solution consists of two new techniques for isolation
drivers and then recovering from their failure.
First, I present an architecture, called Nooks, for isolating drivers from the kernel in a new
protection mechanism called a lightweight kernel protection domain. By executing drivers within a
domain, the kernel is protected from their failure and cannot be corrupted.
Second, I present a mechanism, called shadow drivers, for recovering from device driver fail-
ures. Based on a replica of the driver’s state machine, a shadow driver conceals the driver’s failure
from applications and restores the driver’s internal state to a point where it can process requests as
if it had never failed. Thus, the entire failure and recovery is transparent to applications.
I also show that Nooks functions as a platform for higher-level reliability services. I leverage
the Nooks and shadow driver code to remove another major source of downtime: reboots to update
driver code. This service, dynamic driver update, replaces drivers online, without restarting the OS
or running applications and without changes to driver code. This service requires little additional
code, and demonstrates that Nooks and shadow drivers provide useful services for constructing
additional reliability mechanisms.
I implemented the Nooks architecture, shadow drivers, and dynamic driver update within the
Linux operating system and used them to fault-isolate several device drivers and other kernel ex-
tensions. My results show that Nooks offers a substantial increase in the reliability of operating
systems, catching and quickly recovering from many faults that would otherwise crash the system.
Under a wide range and number of fault conditions, I show that Nooks recovers automatically from
99% of the faults that otherwise cause Linux to crash. I also show that the system can upgrade exist-
ing drivers without rebooting. Finally, I show that the system imposes little performance overhead
for many drivers.
TABLE OF CONTENTS
List of Figures iv
List of Tables vi
Chapter 1: Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Thesis and Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Dissertation Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
Chapter 2: Background and Related Work 6
2.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 System Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Failure Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.4 Properties of OS Failures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.5 Previous Approaches to Building Highly Reliable Systems . . . . . . . . . . . . . 12
2.6 Fault Tolerance in Nooks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
Chapter 3: Device Driver Overview 28
Chapter 4: The Nooks Reliability Layer 31
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.2 Nooks Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.3 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
i
Chapter 5: The Shadow Driver Recovery Mechanism 56
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5.2 Shadow Driver Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.3 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Chapter 6: Evaluation of Nooks and Shadow Drivers 73
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
6.2 Test Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
6.3 System Survival . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.4 Application Survival . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
6.5 Testing Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.6 Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
6.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
Chapter 7: The Dynamic Driver Update Mechanism 98
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
7.2 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
7.3 Design and Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
7.4 Evaluation of Dynamic Driver Update . . . . . . . . . . . . . . . . . . . . . . . . 112
7.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
Chapter 8: Lessons 121
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
8.2 Driver Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
8.3 Driver Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
8.4 Memory Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
8.5 Source Code Availability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
8.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
ii
Chapter 9: Future Work 129
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
9.2 Static Analysis of Drivers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
9.3 Hybrid Static/Run-Time Checking . . . . . . . . . . . . . . . . . . . . . . . . . . 130
9.4 Driver Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
9.5 Driverless Operating Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
9.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
Chapter 10: Conclusions 135
Bibliography 137
iii
LIST OF FIGURES
Figure Number Page
3.1 A sample device driver. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.2 Examples of driver state machines. . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.1 The architecture of Nooks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.2 The Nooks Reliabilty Layer inside the Linux OS. . . . . . . . . . . . . . . . . . . 38
4.3 Protection of the kernel address space. . . . . . . . . . . . . . . . . . . . . . . . . 39
4.4 Trampoline wrappers for invoking multiple implementations of an interface function. 42
4.5 Control flow of driver and kernel wrappers. . . . . . . . . . . . . . . . . . . . . . 43
4.6 Code sharing among wrappers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5.1 A sample shadow driver operating in passive mode. . . . . . . . . . . . . . . . . . 60
5.2 A sample shadow driver operating in active mode. . . . . . . . . . . . . . . . . . . 60
5.3 The Linux operating system with several device drivers and the shadow driver re-
covery subsystem. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.4 The state machine transitions for a sound-card shadow driver . . . . . . . . . . . . 63
6.1 The reduction in system crashes observed using Nooks. . . . . . . . . . . . . . . . 79
6.2 The reduction in non-fatal extension failures observed using Nooks. . . . . . . . . 80
6.3 Results of fault-injection experiments on Linux-SD. . . . . . . . . . . . . . . . . . 88
6.4 Comparative times spent in kernel mode for the Compile-local (VFAT) benchmark. 94
7.1 A shadow driver capturing the old driver’s state. . . . . . . . . . . . . . . . . . . . 105
7.2 Device driver and shadow after an update. . . . . . . . . . . . . . . . . . . . . . . 110
7.3 Application performance on Linux-DDU relative to Linux-Native . . . . . . . . . . 118
iv
7.4 Absolute CPU utilization on Linux-DDU and Linux-Native. . . . . . . . . . . . . . 119
v
LIST OF TABLES
Table Number Page
2.1 Comparison of reliability techniques. . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.1 The number of non-comment lines of source code in Nooks. . . . . . . . . . . . . 38
5.1 The proxying actions of the shadow sound-card driver. . . . . . . . . . . . . . . . 69
5.2 The size and quantity of shadows and the drivers they shadow. . . . . . . . . . . . 70
6.1 The Linux extensions tested. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
6.2 The types of faults injected into extensions. . . . . . . . . . . . . . . . . . . . . . 76
6.3 The applications used for evaluating shadow drivers. . . . . . . . . . . . . . . . . 83
6.4 The observed behavior of applications following the failure of a device driver. . . . 84
6.5 The relative performance of Nooks compared to native Linux. . . . . . . . . . . . 90
7.1 The dynamic driver update process. . . . . . . . . . . . . . . . . . . . . . . . . . 104
7.2 The number of lines of code dynamic driver update. . . . . . . . . . . . . . . . . . 111
7.3 The three classes of shadow drivers and the Linux drivers tested. . . . . . . . . . . 113
7.4 The results of dynamically updating drivers. . . . . . . . . . . . . . . . . . . . . . 114
7.5 The results of compatibility tests. . . . . . . . . . . . . . . . . . . . . . . . . . . 116
7.6 The applications used for evaluating dynamic driver update performance. . . . . . 117
vi
ACKNOWLEDGMENTS
Thank you all!
I would like to thank my advisors, Hank Levy and Brian Bershad, who encouraged me in this
endeavor and taught me how to do research. When I entered graduate school, I thought I knew
everything. From Hank and Brian I learned patience, as research is not an overnight process. I
learned the value of a good question. I learned the value of a good story. And most importantly, I
learned the value of a clear presentation, because ideas are useless if nobody understands them.
I would also like to thank all the students who participated in the Nooks project. Steve Mar-
tin provided early testing. Doug Buxton and Steve Martin implemented our wrapper generator.
Leo Shum implemented sound-card driver support Muthu Annamalai implemented IDE support.
Christophe Augier performed the fault injection testing for the SOSP paper and greatly enhanced
Nooks recovery ability through a scripting language. Damien Martin-Guillerez improved the wrap-
per code and implemented dynamic driver update. Brian Milnes performed performance testing
and optimization, using super-pages to reduce the number of TLB misses. Micah Brodsky and Eric
Kochhar implemented a reliability subsystem in the Windows kernel and provided great feedback
on the assumptions of my system.
I would also like to thank Rob Short at Microsoft, who provided valuable information about
Windows failures and the problem being faced in its development.
Finally, I would like to thank my parents, Alice and Arthur Swift, who allowed me to choose
any career I wanted but are secretly happy that I’ve followed in their footsteps.
vii
DEDICATION
To my family: Suzanne, Jacob and Lily, who encouraged me in my passions, bravely weathered the
resulting absences, and followed me to a land of snow and ice.
viii
1
Chapter 1
INTRODUCTION
This dissertation presents an architecture and mechanism for improving the reliability of com-
modity operating systems by enabling them to tolerate component failures.
Improving reliability is one of the greatest challenges for commodity operating systems. System
failures are commonplace and costly across all domains: in the home, in the server room, and in
embedded systems, where the existence of the OS itself is invisible. At the low end, failures lead to
user frustration and lost sales. At the high end, an hour of downtime from a system failure can lead
to losses in the millions [77].
1.1 Motivation
Three factors motivated my research. First, computer system reliability remains a crucial but un-
solved problem [86, 164]. This problem has been exacerbated by the adoption of commodity
operating systems, designed for best-effort operation, in environments that require high availabil-
ity [187, 20, 4, 56]. While the cost of high-performance computing continues to drop because of
commoditization, the cost of failures (e.g., downtime on a stock exchange or e-commerce server,
or the manpower required to service a help-desk request in an office environment) continues to
rise as our dependence on computers grows. In addition, the growing sector of “unmanaged” sys-
tems, such as digital appliances and consumer devices based on commodity hardware and soft-
ware [102, 194, 180, 184, 179], amplifies the need for reliability.
Second, OS extensions have become increasingly prevalent in commodity systems such as Linux
(where they are called modules [31]) and Windows (where they are called drivers [55]). Extensions
are optional components that reside in the kernel address space and typically communicate with the
kernel through published interfaces. Common extensions include device drivers, file systems, virus
2
detectors, and network protocols. Drivers now account for over 70% of Linux kernel code [45], and
over 35,000 different drivers with over 112,000 versions exist on Windows XP desktops [151, 160].
Unfortunately, most of the programmers writing drivers work for independent hardware vendors and
have significantly less experience in kernel organization and programming than the programmers
that build the operating system itself.
Third, device drivers are a leading cause of operating system failure. In Windows XP, for exam-
ple, drivers cause 85% of reported failures [160]. In Linux, the frequency of coding errors is up to
seven times higher for device drivers than for the rest of the kernel [45]. While the core operating
system kernel can reach high levels of reliability because of longevity and repeated testing, the ex-
tended operating system cannot be tested completely. With tens of thousands of drivers, operating
system vendors cannot even identify them all, let alone test all possible combinations used in the
marketplace. In contemporary systems, any fault in a driver can corrupt vital kernel data, causing
the system to crash.
1.2 Thesis and Contributions
My thesis is that operating system reliability should not be limited by extension reliability. An
operating system should never crash due to a driver failure. Therefore, improving OS reliability
will require systems to become highly tolerant of failures in drivers and other kernel extensions.
Furthermore, the hundreds of millions of existing systems executing tens of thousands of drivers
demand a reliability solution that is both backward compatible and efficient for common device
drivers. Backward compatibility improves the reliability of already deployed systems. Efficiency
avoids the classic trade-off between robustness and performance.
To this end, I built a new operating system subsystem, called Nooks, that allows existing device
drivers and other extensions to execute safely in commodity kernels. Nooks acts as a layer be-
tween extensions and the kernel and provides two key services–isolation and recovery. This thesis
describes the architecture, implementation and performance of Nooks.
Nooks allows the operating system to tolerate driver failures by isolating the OS from device
drivers. With Nooks, a bug in a driver cannot corrupt or otherwise harm the operating system.
Nooks contains driver failures with a new isolation mechanism, called a lightweight kernel pro-
3
tection domain, that is a privileged kernel-mode environment with restricted write access to kernel
memory. Nooks interposes on all kernel-driver communication to track and validate all modifica-
tions to kernel data structures performed by the driver, thereby trapping bugs as they occur and
facilitating subsequent recovery. Unlike previous safe-extension mechanisms, Nooks supports ex-
isting, unmodified drivers and requires few changes to existing operating systems.
When a driver failure occurs, Nooks detects the failure with a combination of hardware and
software checks and triggers automatic recovery. Nooks restores driver operation by unloading and
then reloading the driver. This ensures that the driver services are available to applications after
recovery.
With Nooks’ recovery mechanism, the kernel and applications that do not use a failed driver
continue to execute properly, because they do not depend on the driver’s service. Applications that
depend on it, however, cannot. These failures occur because the driver loses application state when
it restarts, causing applications to receive erroneous results. Most applications are unprepared to
cope with this. Rather, they reflect the conventional failure model: drivers and the operating system
either fail together or not at all.
To remedy this situation, I introduce a new kernel agent, called a shadow driver, that conceals
a driver’s failure from its clients while recovering from the failure. During normal operation, the
shadow tracks the state of the real driver by monitoring all communication between the kernel and
the driver. When a failure occurs, the shadow inserts itself temporarily in place of the failed driver,
servicing requests on its behalf. While shielding the kernel and applications from the failure, the
shadow driver restarts the failed driver and restores it to a state where it can resume processing
requests as if it had never failed.
The two services provided by Nooks and shadow drivers, transparent isolation and recovery,
provide a platform for constructing additional reliability services. I built a third service on top of
these components that enables device drivers to be updated on-line without rebooting the operat-
ing system. This new mechanism, called dynamic driver update, relies upon shadow drivers both
to conceal the update and to transfer the state of the old driver to the new driver automatically,
ensuring that driver requests will continue to execute correctly after the update. Dynamic driver up-
date improves availability by removing the lengthy reboot currently required to apply critical driver
patches.
4
My focus on reliability is not new. The last forty years has produced a substantial amount of
research on improving reliability for kernel extensions by new kernel architectures [51, 212, 71, 97],
new driver architectures [169, 83], user-level extensions [82, 132, 212], new hardware [74, 209], and
type-safe languages [24, 87, 110].
While previous systems have used many of the underlying techniques in Nooks, my system
differs from earlier efforts in three key ways. First, I target an existing set of extensions, device
drivers, for existing commodity operating systems instead of proposing a new extension architecture.
I want today’s drivers to execute on today’s platforms without change if possible. Second, I use C,
a conventional programming language. I do not ask developers to change languages, development
environments, or, most importantly, perspective. Third, unlike earlier safe-extension systems that
were only designed to isolate extensions, Nooks also recovers by restarting drivers and concealing
failures from applications. Nooks does not require modifications to applications to tolerate driver
failures. Overall, I focus on a single and very serious problem – reducing the huge number of
crashes resulting from drivers and other extensions – instead of a more general-purpose but higher-
cost solution.
I implemented a prototype of Nooks in the Linux operating system and experimented with a
variety of device drivers and other kernel extensions, including a file system and a kernel-mode
Web server. Using synthetic fault injection [108], I show that when I inject bugs into running
extensions, Nooks can gracefully recover and restart the extension in 99% of the cases that cause
Linux to crash. Recovery occurs quickly, as compared to a full system reboot, and is unnoticed by
running applications in 98% of tests. The performance overhead is low to moderate, depending on
the frequency of communication between the driver and the kernel. I also show that dynamic driver
update allows transparent, on-line upgrade between a wide range of driver versions. Finally, of the
14 drivers I isolated with Nooks, none require code changes. Of the two kernel extensions, one
required no changes and only 13 lines changed in the other.
Although I built the prototype on the Linux kernel, I expect that the architecture and many
implementation features would port readily to other commodity operating systems, such as BSD,
Solaris, or Windows.
5
1.3 Dissertation Organization
In this dissertation, I describe the architecture and implementation of Nooks and evaluate its ef-
fectiveness and performance cost. Chapter 2 provides background material on reliability, failure
models, and the causes of operating system failure. I also discuss previous work on operating sys-
tem reliability, including techniques for avoiding bugs from OS code and for tolerating faults at
run-time. As this dissertation concerns device drivers, Chapter 3 presents background material on
driver architectures and operations.
Chapter 4 describes the Nooks Reliability Layer, which prevents drivers from corrupting the OS
by isolating drivers from the operating system kernel. After a failure, Nooks recovers by unloading
and reloading the failed driver. I first describe the architecture of Nooks at a high level and then
discuss the details of the Nooks implementation in Linux.
Chapter 5 describes shadow drivers, a replacement recovery mechanism for Nooks that conceals
the failure from applications and the OS. This allows applications that depend on a driver’s service
to execute correctly in the presence of its failure. Again, I provide a high-level view of shadow
driver operation and then discuss the implementation details.
Following the chapters on Nooks and shadow drivers, Chapter 6 analyzes the effectiveness of
Nooks and shadow drivers at (1) preventing drivers from crashing the OS and (2) recovering from
driver failures. I also analyze the performance cost of using Nooks with common drivers and appli-
cations.
Chapter 7 describes dynamic driver update, a mechanism based on shadow drivers to update
device drivers on-line. I present an evaluation showing that dynamic update works for most common
driver updates and imposes a minimal performance overhead.
In Chapter 8 I discuss the lessons learned in implementing Nooks. I evaluate the impact of
operating system architecture on the Nooks architecture by comparing the implementation on Linux
to a partial implementation on Windows. I discuss opportunities for future research in Chapter 9.
Finally, in Chapter 10 I summarize and offer conclusions.
6
Chapter 2
BACKGROUND AND RELATED WORK
The goal of my work is to improve the reliability of commodity operating systems, such as
Linux and Windows. Reliability has been a goal of computer systems designers since the first
days, when hardware failures were orders of magnitude more common then they are today [204,
63]. As hardware reliability has improved, software has become the dominant source of system
failures [91, 150]. My work therefore focuses on improving operating system reliability in the
presence of software failures.
In this section, I present background material on reliable operating systems. First, I define the
terms used to discuss reliable operation. I then discuss software failures and their manifestations.
Finally, I describe previous approaches to building reliable systems and their application to com-
modity operating systems.
2.1 Definitions
Reliability is defined as the expected time of correct execution before a failure and is measured
as the mean time to failure (MTTF) [93]. Availability is the proportion of time that the system is
available to do useful work. Availability is calculated from the MTTF and the duration of outages
(the mean time to repair, or MTTR) as MTTF/(MTTF+MTTR) [93]. My work seeks to improve
reliability by reducing the frequency of operating system failures, and to improve availability by
reducing the time to repair, or recover from a failure.
The events that cause a system to be unreliable can also be precisely defined [93, 10]. A fault is a
flaw in the software or hardware. When this fault is executed or activated, it may corrupt the system
state, leading to an error. If the error causes the system to behave incorrectly, a failure occurs. An
error is termed latent until it causes a failure, at which point it becomes effective.
As an example, a bug in a device driver that corrupts memory is a fault. When it executes and
7
actually corrupts memory, the resulting corruption is an error. Finally, a failure occurs when the
kernel reads the corrupt memory and crashes the system.
2.2 System Models
There exist a wide variety of operating systems that target different environments. For example,
the QNX operating system targets real-time services on embedded devices [104]. Unicos targets
high-performance applications on supercomputers [29], and the Tandem Guardian operating system
targets application with extreme reliability needs, such as stock exchanges [128]. These special-
purpose operating systems are optimized for the unique properties of their operating environment
and workload, for example by trading peak performance for predictable scheduling.
In contrast, commodity operating systems, such as Linux [31] and Windows [173], are sold in
high volumes and used in a wide variety of environments. For example, in addition to use on desk-
tops and in data centers, Linux has been used on wristwatches [156] and in planetary exploration [4].
As a result, commodity operating systems cannot be over-optimized for a single target environment.
Instead, they are designed with a variety of competing goals, such as cost, size, peak performance,
features, and time-to-market. Furthermore, the huge installed base of these operating systems forces
compatibility with existing code to be a priority. As a result, any attempt to improve the reliability
of commodity operating systems must take into account the wide variety of demands placed on the
system.
2.3 Failure Models
Failures occur when a fault is activated and causes an effective error. The circumstances of when
and how a fault causes a failure, and the failure’s behavior, determine what a system must do to
tolerate the failure. As a result, designers make many assumptions about how failures occur in order
to build a cost-effective system that prioritizes frequent and easy-to-handle failures. For example,
most systems are built with the assumption that memory will not become corrupted silently. As a
result, these systems avoid the cost of error checking every memory operation. If silent memory
corruption occurs, the system may cease functioning. The same system, though, may also assume
that a network can corrupt data in transmission and, as a result, checksums packet contents to detect
8
corruption. Thus, the assumptions about what failure occurs and their frequency drive system design
in terms of what mechanisms are used to detect and tolerate failures.
A system’s assumptions about failures are called its failure model. This model specifies any
dependency that a system has on the properties of failures. In the context of software reliability,
failure models commonly specify two major properties of failures: how the system behaves when it
fails and how failures occur. How the system behaves determines what mechanisms are required to
detect failures. How failures occur determines whether a failure can be masked from other system
components.
A failure model may also incorporate assumptions about the system’s response to a component
failure. Rarely do entire systems fail at once. Generally, a single component fails, which if not
handled, leads to a system failure. To simplify failure handling, a system designer may assume
that failures are contained to a single component and detected before any other component has been
corrupted. In contrast to the previous assumptions, this is an assumption about the system and not
the failure, and must therefore be enforced by the implementation.
2.3.1 How Failures Behave
The failure behavior of a component determines how a system contains and detects failures. For
example, a system can detect component crashes with the processor’s exception mechanism. To
contain a component that fails by producing incorrect outputs, the system must check all outputs
for correctness before other components observe the output and become corrupted. In contrast, if a
component fails by stopping, then no work may be needed to contain its failure.
We can categorize failures by their behavior in order to detect and tolerate different behaviors
with different mechanisms. Previous studies have identified five major categories of failure behav-
ior [54]:
1. Timing failures occur when a component violates timing constraints.
2. Output or response failures occur when a component outputs an incorrect value.
3. Omission failures occur when a component fails to produce an expected output.
4. Crash failures occur when the component stops producing any outputs.
5. Byzantine or arbitrary failures occur when any other behavior, including malicious behavior,
9
occurs.
Each failure category requires its own detection and containment mechanisms. For example, omis-
sion failures require that the system predict the outputs of a component to detect when one is miss-
ing, while timing failures require timers that detect when an output is delayed.
For simplicity, reliable systems commonly assume that only a subset of these behaviors occurs
and remove from consideration the other failure modes. For example, Backdoors [27] assumes that
only timing and crash failures occur. Fortunately, many failures are of these types, although the more
general arbitrary failure has been studied in context of some byzantine-tolerant research operating
systems [177, 37].
2.3.2 How Failures Occur
How a fault is activated and causes a failure determines what strategies are possible for tolerating
the failure. A fault that is activated by a single, regularly occurring event is called a deterministic
failure, or a bohrbug [91]. Any time that event occurs, the failure occurs. As a result, a recovery
strategy for deterministic failures must ensure that the triggering event does not occur again during
recovery. If the triggering event is a specific service request, that request must be canceled, because
it can never be completely correctly.
In contrast, a fault that is activated by a combination of events that occur irregularly is called
a transient failure, or a heisenbug. These failures are triggered by a combination of inputs, such
as the exact interleaving of requests and other properties of the system environment that do not
recur often [91]. As a result, the recovery strategy need not prevent the failure from recurring–it is
probabilistically unlikely to happen. Requests being processed at the time of failure can be retried
automatically, which conceals the failure from the component’s clients. Empirical evidence suggests
that many failures are transient [91] and as a result many fault-tolerant systems ignore deterministic
failures. For example, the Phoenix project [15], the TARGON/32 fault-tolerant Unix system [28]
and Shah et al’s work on fault-tolerant databases [183] all assume that all failures are transient.
10
2.3.3 System Response to Failure
When a component failure occurs, the system may translate the failure from one category to another.
For example, Windows XP crashes when it detects that a device driver passes a bad parameter to a
kernel function. Thus, the OS translates an output failure of a component into a crash failure of the
system. Failure translation allows higher levels of software to process a single category of failure
rather than manage the variety of component failures that may occur.
Many fault-tolerant systems take this approach with the fail-stop processor model [175], whereby
the system halts a component when it detects a failure. By halting, the component cannot further
spread corruption. A system must meet three conditions to be fail-stop:
1. Halt on failure: the processor halts before performing an erroneous state transformation.
2. Failure status: the failure of a processor can be detected.
3. Stable storage: the processor state can be separated into volatile storage, which is lost after a
failure, and stable storage, which is preserved, unaffected by the failure.
The benefit of the fail-stop model is that it greatly simplifies construction of a reliable system: it
is simple to distinguish between correct and failed states and memory is similarly partitioned into
correct and corrupted storage. Thus, the system and its components are only concerned with these
two states and not with other failure modes. Many fault-tolerant systems, such as Phoenix [15]
assume that failures are fail stop. Others, such as Hive [39] incorporate failure detectors that translate
output failures into fail-stop failures.
The drawback of the fail-stop model is that it is difficult to achieve in practice and does not
adequately capture all the properties of the system, such as intermittent failures [8]. Furthermore,
without strong isolation that contains faults to a single component and good failure detectors, it
can be difficult to ensure fail-stop behavior. For example, a study by Chandra and Chen shows
that application failures are often detected after corrupt data is written to disk [38], violating the
fail-stop conditions. Additional error checking internal to the application may be able to reduce the
frequency of the violations.
11
2.3.4 Summary
How failures occur and how failures behave determine the possible strategies for tolerating failures.
If failures are transient, then a recovery strategy may retry operations in progress at the time of
failure. In contrast, if failures are deterministic, care must be taken to avoid repeating the triggering
event. Similarly, if failures have easily detected behavior, such as crashes, or are translated into
fail-stop failures, then the system can avoid managing complex error states.
2.4 Properties of OS Failures
Within the OS, many failures can be traced to operating system extensions. For example, recent
data on the Windows NT family of operating systems from Microsoft implicates operating system
extensions as the major source of failures, particularly as the core operating system itself becomes
more reliable [150]. According to this study, the core operating system in Windows NT 4 was
responsible for 43% of system failures, while drivers and other kernel extensions were responsible
for 32%. In Windows 2000, the follow-on to NT 4, the core operating system was responsible
for only 2% of failures, while drivers grew to 38% (the remaining failures were caused by faulty
hardware and system configuration errors). More recently, drivers caused over 85% of Windows
XP crashes [160]. Hardware failures remain a significant, but not dominant, source of system
crashes: for Windows Server 2003, 10% of failures are due to hardware problems [140]. However,
approximately half of the crashes from hardware faults manifest as driver failures. These driver
failures could be prevented with a “hardened” driver that is able to mask the hardware fault [9].
While there are no published crash-reports for the Linux operating system, an automated study
of the Linux 2.4 kernel source code showed similar trends. Device drivers, which comprised 70%
of the source code, had between two and seven times the frequency of coding errors (errors per line
of code) as other kernel code [45]. Thus, drivers are likely a major source of failures in Linux as
well.
Many studies of operating system failures have shown that these failures are often transient, i.e.,
they do not recur when the same operation is repeated [91, 92]. One reason is that kernel code
is highly concurrent and may run simultaneously on multiple threads or processors. As a result,
the exact ordering of events changes frequently. Other frequent changes in the kernel environment,
12
such as the availability of memory and the arrival of interrupts, can also cause transient failures. In
addition, the dominant manifestation of faults in operating systems is memory corruption [191, 101].
Thus, changes in memory-allocation patterns may also cause failures to be transient.
2.5 Previous Approaches to Building Highly Reliable Systems
Due to the need for reliable systems, many approaches have been tried in the past. At a high
level, past systems have improved reliability with two general techniques: fault avoidance and fault
tolerance. Fault avoidance seeks to avoid executing faulty code, whereas fault tolerance seeks to
continue executing after a fault is activated.
2.5.1 Fault Avoidance
The goal of fault avoidance is to improve reliability by executing only correct code. The strategy can
applied at all points during the lifecycle of software: during the design and coding process, when
it is called fault prevention, after code is written, called fault removal, or at run time, called fault
work-arounds.
Fault Prevention
The goal of fault prevention is to ensure that faults never make it into the source code. There
are two major approaches to fault prevention: software engineering techniques and programming
languages. Software engineering specifies how to write software. For example, modular software
structure reduces faults by clarifying the interfaces between components [162, 163]. More recently,
aspect-oriented software development (AOSD) systems promise to reduce software faults in operat-
ing systems by avoiding common cut-and-paste errors [47]. Software engineering processes, such
as RTCA DO178B for aviation software [111] and the Capability Maturity Model’s Level 5 [165],
address not only coding but also planning, documentation, and testing practices. These processes
seek to turn the craft of programming into a rigorous engineering discipline.
Programming languages prevent faults by preventing compilation of faulty code. Type-safe pro-
gramming languages and run-time systems used for operating systems, such as Modula 2+ [141],
Modula 3 [141, 24], Java [87], Cyclone [114], C] [110], Vault [58], and Ivy [34], prevent unsafe
13
memory accesses and thereby avoid the major cause of failures. In addition to checking memory
accesses, languages can provide compile-time checking of other properties. For example, the Vault
language provides an extended type system that enforces ordering and data-flow constraints, such as
the order of operations on a network socket [58]. To date, however, OS suppliers have been unwill-
ing to implement system code in type-safe, high-level languages. Moreover, the type-safe language
approach makes it impossible to leverage the enormous existing code base. Recent approaches
that add type-safety to C, such as CCured, remove the cost of porting but increase run-time costs
substantially because of the need to track the types of pointers [50].
Other language-based approaches seek to improve reliability by automating error-prone man-
ual tasks. Aspect-oriented software development (AOSD) automates the application of changes
that cut across a large body of code [118] and has been applied to operating system kernels [47].
Tarantula applies AOSD to automate changes to the kernel-driver interface [127], while c4 uses
aspect-oriented techniques to create a new programming model for cross-cutting kernel patches [80].
Compared to other language-based techniques, these approaches have the benefit that they are fully
compatible with existing code in that they apply changes to code written in C.
Domain-specific languages can similarly improve reliability by simplifying complex and error-
prone code. The Bossa project, for example, provides a language and run-time system for writing
Linux schedulers that remove much of the complexity in current schedulers [30]. The Devil project
simplifies writing driver code that interfaces with devices [142]. In most operating systems, the
device interface is specified in a document describing the state machine of the device and the orga-
nization of registers exported by the device for communication. Based on this description, the driver
writer must then implement code in C or another low-level language that interacts with the device.
With Devil, a device vendor specifies the device-software interface in a domain-specific language.
The Devil compiler then uses that specification to generate an API (i.e., C-language stubs) for the
device. Driver writers call these functions to access the device. Devil prevents many of the bugs as-
sociated with drivers by abstracting away the complexities of communicating through I/O ports and
memory-mapped device registers. This approach requires additional work by device manufacturers
to produce the specification, but reduces the effort of driver programmers. However, it requires
nearly complete rewriting of existing drivers, and hence is only applicable to new drivers. Devil
complements the other language techniques, in that it addresses the hardware interface instead of
14
the system interface to drivers.
While fault prevention is important to the writing of new code, it is of limited use for the millions
of lines of existing code in operating systems. Fault removal is a more appropriate technique for
this code.
Fault Removal
In contrast to fault prevention, which seeks to prevent faults from being written, fault removal seeks
to detect faults in existing code. There are both process-oriented approaches to fault removal, such as
code reviews [76], and technology approaches, such as static analysis. While programming practices
are important and can have a huge impact on program quality [165], technology for automatically
finding bugs is of more interest given the huge base of existing operating system code. For example,
Microsoft Windows Server 2003 has over 50 million lines of code [193].
There are two major approaches to analyzing source code for faults: theorem proving and model
checking [208]. Theorem proving works by converting the program code into logic statements that
are then used to prove that the program never enters an illegal state. In contrast, model checking
enumerates the program states along possible execution paths to search for an illegal state. Theorem
proving generally scales better, because it manipulates formulas rather than program states. How-
ever, theorem provers often require human input to guide their execution towards a proof [208]. For
this reason, model checkers have proven more popular as tools for a general audience [126].
Model checkers evaluate an abstraction of the code along all paths to ensure that it meets certain
invariants, e.g., that memory that has been released is never accessed again. The output of a model
checker may be: (1) an example path through the code that violates an invariant, (2) termination with
the guarantee that the invariants are never violated, or (3) continued execution as the model checker
works on ever-longer paths. To prove correctness, a model checker must check every possible state
of a program, which is exponential in program size. Hence, model checkers have difficulty scaling
because they must exhaustively enumerate the state space of a program. However, model checking
can find bugs (but not prove correctness) by checking a much smaller number of states [211]. For
example, Microsoft’s PREfix and PREfast tools, which create a simplified abstraction of the source
code to speed checking, do not prove that code is correct but are nonetheless useful for finding
15
common programming errors [126].
Traditionally, model checking could not be applied to program code because of the complexity
of evaluating the code directly. Instead, a programmer had to translate the code into a specification
language such as Promela [67, 106], which is then checked. Recent language designs, such as
ESP [124] and Spec] [18], allow model checking of executable code. The benefit of this approach
is that no separate specification of the code is needed; instead the code itself is directly checked.
The drawback is that it requires writing code in a new language and hence is not compatible with
existing code or programming practices.
Several recent systems automatically compute a model from the existing source code, remov-
ing the need for new languages. The SLAM project from Microsoft creates binary programs from
device drivers that can be checked [14]. Both SLAM and the similar Blast project [100] improve
performance by only refining the models with additional detail when necessary. In addition, recent
work has shown that it is possible to directly model check C code by executing it in a virtual ma-
chine [153, 152]. The virtual machine creates snapshots of each program state. However, model
checking by direct execution is often slow, and it is difficult to isolate the state of the component
being tested from the rest of the system [153]. As a result, this technique is not yet widely used.
Recently, model checking and theorem proving tools have been applied to existing operating
system code [68, 13, 100]. The output of these tools is a list of sites in or paths through the code
that may be faulty. However, finding bugs does not necessarily translate directly into more reliable
software. First, even though a small fraction of faults cause the majority of failures [1], tools are
also unable to determine statically whether a fault is a significant cause of failure [70]. Second,
a programmer must still review the faults and manually correct the code–current tools cannot fix
the code themselves [70]. Microsoft, for example, runs static analysis tools regularly and forces
developers to fix the bugs they find [154]. However, Microsoft cannot force third-party developers
to run the tools over extension code. Coverity runs similar tools over the Linux source tree and posts
bug reports to the Linux Kernel Mailing List [53]. Again, Coverity only examines the code checked
into the Linux kernel source tree. Many kernel extensions are maintained separately and released as
patches to the kernel, and hence do not benefit from the analysis [138]. As a result, these tools often
help the base operating system become more reliable but have less impact on extension reliability.
16
Fault Work-Arounds
When faulty code has been deployed and the resulting failures cannot be tolerated, fault work-
arounds are the remaining possibility for attaining high reliability. Work-arounds prevent unsafe
inputs from reaching a system. For example, a firewall is a fault work-around technique that prevents
malicious requests from arriving at a network service.
In general, it is impossible to predict in advance the exact inputs that must be filtered to avoid
failures [199, 134]. Thus, work-arounds are an ad-hoc technique for preventing failures with known
causes when the software cannot be fixed. Given the wide variety of environments and inputs to
commodity operating systems, this approach is therefore impractical as a general-purpose strategy
for improving OS reliability.
Obtaining Specifications
A common problem for all fault-avoidance tools is the definition of correctness. Most software
is written without a formal specification of its behavior. Indeed, writing a formal specification is
prohibitively expensive: although the space shuttle control software contains only 425,000 lines
of code, the specifications constitute 40,000 pages [79]. Therefore, many static analysis tools ver-
ify partial specifications, e.g., proving that invariants are not violated instead of complete correct-
ness. The Static Driver Verifier from Microsoft [14], the Vault programming language for device
drivers [58], and the CMC tool for checking network protocols and file systems [153, 152, 211] use
this approach.
An existing body of code can be another source of specifications. If there are many implementa-
tions of an interface, then the code that deviates in its usage of the interface is likely to be incorrect.
For example, if 99% of the implementers of an interface lock a parameter before accessing it, the
remaining 1% of implementations that do not are probably faulty. This approach has been applied
to the Linux kernel to identify both extension interface rules and violations of those rules [69].
A third approach to obtaining specifications for system software is to write them by hand. For-
mally verified operating system code [67, 198] and high-assurance systems [26] take this approach.
However, only subsystems or small operating systems have been verified to date. Given the size of
commodity operating systems, it is impractical to write a complete specification by hand.
17
Fault Avoidance Summary
Fault avoidance improves reliability by not executing faults. The choice between prevention, re-
moval, and work around, depends on a projects stage of development. Fault prevention applies to
systems in the design stage, when the system is being architected and implemented. Fault removal
applies to existing code bases. Fault work-arounds apply to deployed systems and provide a stopgap
solution until the system can be patched.
2.5.2 Fault tolerance
The previous section described methods for avoiding bugs by preventing faulty code from being
executed. In contrast, fault tolerance seeks to continue correct execution in the presence of faulty
code.
The core technique for tolerating failures is redundancy, so that a single failed computation does
not terminate execution. The redundancy may come physically, by performing the computation on
multiple processors, temporally, by re-executing the computation at a later time, or logically, by
executing on a different copy of the data. A system may also use multiple implementations of a
component to increase the probability that some computation will complete successfully. The core
assumption behind redundancy is failure independence: the different implementations or executions
will not all fail simultaneously.
At the program design level, there are two major categories of fault-tolerant designs: multi-
version and single-version [195]. In multi-version designs, code components are implemented mul-
tiple times, providing design diversity and lessening the probability that all versions fail on the same
inputs. When one version does fail, the other versions should (hopefully) continue to execute cor-
rectly. Single-version designs, in contrast, execute only one version of the code and instead rely on
additional code to detect and recover from failures.
Three major multi-version approaches have been developed to provide both diversity and re-
dundancy: recovery blocks [170], N-version programming [41], and N self-checking [125]. While
the details differ, all three techniques require that programmers implement multiple versions of a
component and a mechanism, either a checker or a voter, to select the correct output. Even with
diversity, though, experimental evidence suggests that the different versions of a component often
18
share failure modes, because the difficulty of producing the correct output varies widely across the
space of inputs. Thus, different implementations are likely to fail on the same difficult inputs [121].
Implementing multiple versions of a program is time-consuming and expensive. The BASE
technique reduces this cost by allowing off-the-shelf implementations of standard network services,
such as a file server, to be configured into an N-version system [171]. An abstraction layer smoothes
small differences between implementations, such as the granularity of timestamps and message or-
dering. This approach is limited to applications with well-specified behavior and clean, documented
interfaces. As a result, it is not applicable to whole operating systems, which meet neither condition.
Due to the cost of multi-version approaches, most fault-tolerant systems opt for single-version
redundancy and forgo the benefits of design diversity. Single-version fault tolerance depends on
error detection, to notify the system that an error has occurred, and recovery, to transform the
incorrect system state into a correct system state [10]. In addition, fault tolerant systems generally
rely on isolation or containment to limit the scope of damage from a fault and to prevent errors from
becoming failures [59]. In the next sections I further discuss isolation, error detection, and recovery
as it applies to operating systems.
Isolation
Isolation, which confines errors to a single component, has long been recognized as the key to fault
tolerance because it allows components to fail independently [59, 91].
The coarsest granularity for isolation is the entire computer. Many commercial fault-tolerant
systems, such as Stratus Integrity systems [113] and Tandem NonStop systems [19, 22], execute
each program on multiple processors. If one processor fails, the others continue uncorrupted. There
have been many research systems that take the same approach [28, 11, 33, 177]. The primary benefit
of isolating at the machine level is tolerance of hardware failures: when one machine fails, another
can continue to provide service. The drawback, however, is the substantial cost and complexity of
purchasing and managing multiple computers.
Virtual machine technologies [39, 43, 190, 207, 17] are an alternative solution to isolating unreli-
able code. They promise the software reliability benefits of multiple physical computers but without
the associated costs. In addition, virtual machines reduce the amount of code that can crash the
19
whole machine by running the operating system in user mode and only the virtual machine monitor
in privileged mode. Virtualization techniques typically run several entire operating systems on top
of a virtual machine, so faulty extensions in one operating system cause only that OS instance and
its applications to fail. While applications can be partitioned among virtual machines to limit the
scope of failure, doing so removes the benefits of sharing within an operating system, such as fast
IPC and intelligent scheduling.
However, if the extension executes inside the virtual machine monitor, such as device drivers for
physical devices, a driver fault causes all virtual machines and their applications to fail. Executing
extensions in their own virtual machines solves this problem [83, 130, 72], but raises compatibility
problems. All communication between applications and drivers must then transit an IPC channel.
If the contents of driver requests are unknown, they cannot be passed automatically between virtual
machines.
Within a single operating system, user-mode processes provide isolation for unreliable code [186].
Traditionally, operating systems isolate themselves from applications by executing applications in
a separate user-mode address space. Several systems have moved a subset of device drivers into
user-mode processes to simplify development and improve reliability [109, 117, 146]. Microker-
nels [210, 132, 212] and their derivatives [71, 81, 96] extend the process abstraction to apply to
other kernel extensions as well. Extensions execute in separate user-mode address spaces that in-
teract with the OS through a kernel communication service, such as remote procedure call [23].
Therefore, the failure of an extension within an address space does not necessarily crash the system.
Despite much research in fast inter-process communication (IPC) [23, 132], the reliance on separate
address spaces raises performance concerns that have prevented adoption in commodity systems.
An attempt to execute drivers in user mode in the Fluke microkernel doubled the cost of device
use [201]. In addition, there is the high price of porting or reimplementing required to move exten-
sions into user mode. Microkernel/monolithic hybrids, such as L4Linux [97], provide a transition
path to microkernels by executing a monolithic kernel on top of a microkernel.
Protecting individual pages within an address space provides fine-grained isolation and effi-
cient sharing, because data structures need not be marshaled to be passed between protection do-
mains. Page-level protection has been used in single-address-space operating systems [40] and to
isolate specific components or data from corruption, e.g., in a database [192] or in the file system
20
cache [157].
There have been several major hardware approaches to providing sub-page memory isolation.
Capability-based architectures [107, 159, 131] and ring and segment architectures [112, 174] both
enable fine-grained protection. With this hardware, the OS is extended not by linking code into the
kernel but by adding new privileged subsystems that exist in new domains or segments. Several
projects also provide mechanisms for new extensions, written to a new programming model, to
execute safely with segment-based protection [44, 185] on x86 hardware. In practice, segmented
architectures have often been difficult to program and plagued by poor performance [48]. These
problems have been addressed recently with the Mondrian protection model that uses segments
only for protection and not for addressing [209].
To remove the reliance on new (and unavailable) or existing (and slow) hardware memory pro-
tection, software fault isolation (SFI) rewrites binary executables to isolate extension code [205].
The rewriting tool inserts checks before all memory accesses and branches to verify that the target
address is valid. The system detects a failure when the extension references memory or branches
outside safe regions. While SFI provides isolation without hardware support, it requires contiguous
memory regions for efficient address checking, making it inappropriate for existing extensions that
access widely dispersed regions of the kernel address space.
Despite decades of research on isolation, commodity operating systems still rely on user-mode
processes as their only isolation mechanism. However, processes have proved too expensive to use
in many situations, and reliability suffers as a result. Too often, new isolation mechanisms designed
to reduce this cost do not consider recovery or backward compatibility, and hence have found little
use. As a result, there is still a need for lightweight isolation mechanisms for containing accidental
software failures.
Failure detection
Failure detection is a critical piece of reliable execution because it triggers recovery. Prompt failure
detection allows the construction of simpler programming models, such as fail-stop processors, from
complex failure modes. The difficulty of detecting failures depends on the fault model. For example,
timing failures can be detected with a simple timeout mechanism, while detecting output failures
21
may require complete knowledge of the expected output of a component.
Commonly, operating system kernels are implemented with the assumption that faults in kernel
code have been avoided and need not be tolerated. Failure-detection code instead targets user-mode
processes and external devices rather than kernel components. For example, system calls check
their parameters for correctness to ensure that a bad parameter does not crash or compromise the
kernel [186]. Network protocols rely on timeouts to detect dropped packets [168]. Some kernels,
including Linux [31], also do limited exception handling in kernel mode to continue execution in
the presence of non-fatal processor exceptions.
Operating systems that target high reliability add additional checks to ensure that the kernel itself
is functioning. For example, the kernel may monitor progress of its components via counters and
sensors [27, 202]. In a distributed system, a remote machine or independent hardware with access to
the kernel’s address space may do the monitoring [27], allowing remote detection of kernel failures
at low latency. Failures can also be detected probabilistically based on observed deviations from
past behavior [42].
When commodity software is used for applications that need high reliability, wrappers inter-
posed between the commodity component and the high-reliability system can provide error detec-
tion. For example, Mafalda places wrappers around a commodity microkernel to both check the
correctness of the microkernel’s actions and recover after a failure [73]. The Healers system detects
bad parameters passed to standard library routines to convert crash failures (which may occur when
the library operates on bad inputs) into fail-stop failures [78]. Other work has focused on determin-
ing where to place wrappers to limit the scope of corruption [115]. One difficulty with wrappers is
detecting invalid outputs. Generally, unless the correct output is known (via a redundant computa-
tion), then only a subset of invalid outputs can be detected. In the context of device drivers, which
interact with agents outside the computer, it is impossible to detect whether data was corrupted as it
passed from the device through the driver.
Failure detection accuracy fundamentally limits the availability of a system. When failures are
not detected promptly, they persist longer and lower availability [42]. However, simplistic failure
detectors, such as exception handlers and timers that are simple to implement and inexpensive to
execute, may detect sufficiently many failures to improve the reliability of commodity operating
systems.
22
Recovery
The goal of recovery is to restore the system to correct operation after a failure. The database
community has long recognized the importance of recovery and relies on transactions to prevent
data corruption and to allow applications to manage failure [90]. Recently, the need for failure
recovery has moved from specialized applications to the more general arena of commodity systems
and applications [164].
We can classify recovery strategies by the state of the system after recovery and by their trans-
parency to applications and system components. A recovery strategy may move the state of a system
backwards, to a correct state that existed before the failure, or forwards, to a new state [10]. Fur-
thermore, recovery strategies can either be concealing, in that they hide failure and recovery from
applications, or revealing, in that they notify applications of a failure and depend on them to partic-
ipate in the recovery process.
A general approach to recovery is to run application or operating system replicas on two ma-
chines, a primary and a backup. All inputs to the primary are mirrored to the backup. After a
failure of the primary, the backup machine takes over to provide service. The replication can be
performed by the hardware [113], at the hardware-software interface [33], at the system call inter-
face [11, 28, 32], or at a message passing or application interface [19]. For example, process pairs
execute each application on two machines. Before processing a request, the primary forwards the
request to the backup. If the primary fails to complete the request, the backup is then responsible for
recovering, possibly by retrying the failed request [19]. A problem with primary-backup systems
is that extensive failure detection is still required. When the backup and primary processes pro-
duce different results, it is impossible to tell which is correct. Triple-modular redundancy (TMR)
systems execute three copies of the code to avoid the need for extra failure detection and recovery
code [139]. If an output from one replica differs from the other two, it is assumed to be incorrect
and that replica is marked as failed. Thus, no extra checks or recovery code are required, as there
are still two correctly functioning processes [22].
Primary-backup systems generally take a backwards recovery approach. The backup process
begins executing at a point before the error occurred, or, if no error occurred on the backup, at
the exact point of failure. These systems can either conceal or reveal errors from applications.
23
The Tandem NonStop Advanced system, for example, can either expose failures and subsequent
recovery to application using transactions [90] or transparently conceal failure and let the backup
applications execute without knowledge of the failure [22].
On a single machine, transactions provide a revealing mechanism for backwards recovery. Trans-
actions allow the state of a system to be rolled back to a point before the failure occurred [90]. Upon
detecting a failure, applications explicitly abort transactions, which rolls back the system state.
While most commonly used in databases and other applications with persistent state, Vino [182]
and QuickSilver [176] applied transactions within the OS to improve reliability. In some cases, such
as the file system, the approach worked well, while in others it proved awkward and slow [176]. The
revealing strategy of transactions, though, makes them difficult to integrate into existing systems,
which are unprepared for the independent failure of system components.
Another, simpler, revealing strategy is to restart a failed component and, if necessary, its clients.
This strategy, termed “micro-reboots,” has been proposed as a general strategy for building high-
availability software because it reuses initialization code to recover from a failure rather than relying
on separate recovery code [36]. In these systems, the burden of recovery rests on the clients of the
restarted component, which must decide what steps to take to continue executing. In the context of
operating system component failures, few existing applications do this, since these failures typically
crash the system [36]. Those applications that do not may share the fate of the failed component
and must be restarted as well.
A common approach to concealing failures is to checkpoint system state periodically and roll-
back to a checkpoint when a failure occurs [167]. This is a backwards recovery strategy, because
the system returns to a previous state. Logging is often used to record state changes between check-
points [16, 136, 149, 11, 28, 172]. These systems conceal errors from applications by transparently
restarting applications from the checkpoint after a failure and then replaying logged requests. Recent
work has shown that this approach is limited when recovering from application faults: applications
often become corrupted before they fail; hence, their checkpoints and persistent storage may also be
corrupted [38, 135].
The fundamental recovery problem for commodity operating systems is keeping applications
running. Revealing strategies are unworkable, because they depend on application participation
to recover completely. Similarly, forward recovery is also inappropriate, because the state of the
24
system advances beyond what applications may have expected. As a result, improving the reliability
of commodity operating systems demands a concealing backwards recovery mechanism.
Fault Tolerance Summary
Tolerating faults requires isolating faulty code, detecting failures, and recovering from a failure.
Unlike highly reliable systems, today’s commodity operating systems are built with the assumption
that the kernel and its extensions cannot fail, and as a result lack all three elements. The challenge for
existing operating systems is to find efficient and inexpensive mechanisms to provide these elements
while maintaining compatibility with existing code.
2.5.3 Summary
Systems can be made reliable by either avoiding faults and the subsequent failures or by tolerating
failures. The choice of strategy depends largely on the goals and properties of the system. Often,
a hybrid approach is the most appropriate, for example preventing most failures with compile-time
checking and tolerating the remaining ones with isolation and recovery techniques.
In safety critical applications, the code must produce the correct result. As a result, fault-tolerant
strategies relying on retrying or executing multiple versions of a component are not sufficient if none
of the versions produce a correct output. Thus, applications demanding extreme reliability, such as
flight-control systems, rely on programming languages, verification tools, and strict engineering
practices to avoid faults during the development process [111, 165, 26].
An important limitation of fault-avoidance techniques is that they can only verify what can be
specified. For operating system extensions, this has been the interface between the kernel and the
extension. Thus, static analysis tools for drivers only address the correctness of software interfaces
and not the interface to the device hardware. For example, a driver may obey the kernel interface
but pass bad commands to the device, causing the device to perform the wrong action. Hence, static
analysis tools can replace or augment some fault-tolerance techniques, but cannot obviate the need
for fault-tolerant systems as a whole.
25
2.6 Fault Tolerance in Nooks
My work aims to improve the reliability of commodity operating systems with as little mechanism
as possible. I therefore focus on the dominant cause of failures, faulty device drivers. As these
drivers are not under the control of any single entity, I take a fault-tolerance approach that allows
operating systems and applications to continue executing when drivers fail. The paramount goal is
transparency to drivers and applications, neither of which should be modified.
2.6.1 Nooks’ Failure Model
In designing Nooks, I made several assumptions about driver failures. First, I assume that most
driver failures are transient. Thus, a reasonable recovery strategy is to reload the failed driver and
retry requests in progress at the time of failure, because the failure is unlikely to recur. With this
mechanism, it is possible to conceal recovery, because all requests are eventually processed by
the driver. Based on anecdotal evidence, I believe this to be a reasonable assumption. However,
if driver failures are deterministic, concealing the recovery process from applications may not be
possible because requests that trigger the failure can never be successfully completed. Recovery is
still possible, but applications must participate by handing failed requests..
Second, I assume that most driver failures are crash failures, output failures, or timing failures. I
further assume that these failures are fail-stop and can be detected and stopped before the kernel or
other drivers are corrupted. I assume that drivers are not malicious and do not suffer from arbitrary
failures. Thus, Nooks includes services to detect driver failures and prevent a failed driver from
corrupting other system components. The ability of Nooks to make driver failures fail-stop depends
both on the quality of isolation and on the state maintained by the device. For example, if a driver
corrupts persistent device state before the failure is stopped, then the stable-storage property is
violated and Nooks cannot recover. As a result, Nooks must detect failures before persistent device
state is corrupted. If this is not possible, then additional recovery strategies, beyond Nooks’ current
abilities, may be required to repair corrupt state as part of recovery.
26
2.6.2 Nooks’ Design
Nooks uses lightweight kernel protection domains to isolate drivers. These domains are in the kernel
address space, allowing fast sharing of read-only data, and rely on page-level protection to support
efficient protected access to non-contiguous data regions. I show that this is sufficient to isolate
unmodified device drivers and other kernel extensions.
Nooks detects driver failures with simple mechanisms. The kernel exception handlers notify
Nooks when a driver generates an exception, such as a protection violation fault. Nooks places
wrappers around drivers to detect certain bad outputs, such as pointers to invalid addresses, and
relies on timers to detect timing failures when a driver is not making progress. Finally, Nooks
detects an output failure when a driver consumes too many kernel resources.
Shadow drivers, my recovery mechanism for Nooks, provides transparent backwards recovery
by concealing failure from applications while resetting the driver to its state before the failure.
Shadow drivers are similar to primary-backup systems in that they replicate all communication
between the kernel and device driver (the primary), sending copies to the shadow driver (the backup).
If the driver fails, the shadow takes over temporarily until the driver recovers. However, shadows
differ from typical replication schemes in several ways. First, because my goal is to tolerate only
driver failures, not hardware failures, both the shadow and the “real” driver run on the same machine.
Second, and more importantly, the shadow is not a replica of the device driver: it implements only
the services needed to manage recovery of the failed driver and to shield applications from the
recovery. For this reason, the shadow can be much simpler than the driver it shadows.
Shadow drivers log and replay requests similarly to checkpoint/replay recovery systems. How-
ever, shadow drivers rely on the shared interface to drivers to log only those inputs that impact
driver behavior. Compared to checkpointing the complete driver state or logging all inputs, the
shadow driver approach reduces the amount of state in the log. Hence, the likelihood of storing
corrupt state, which is possible with full logging or checkpointing, is also reduced.
2.7 Summary
Table 2.1 lists techniques for improving operating system reliability, the approach they take, and
the changes required to use the approach. Nooks is the only fault-tolerance technique that provides
27
Table 2.1: Comparison of reliability techniques.
Technique Provides Approach RequiredChanges toApplicationsand CriticalComponents
Software engineering N/A avoidance OS, extensions,programmers
Languages Isolation / errordetection
avoidance OS, extensions,Programmers
Static analysis Isolation avoidance noneRedundant hardware Isolation / error
detection /recovery
tolerance HW, OS
Virtual machines Isolation tolerance noneOS Structure Isolation tolerance OS, extensionsCapabilities / ring-segment Isolation tolerance HW, OS,
extensionsMondrian Memory Isolation tolerance HW, OSCheckpoint/Log Recovery tolerance noneTransactions Isolation /
recoverytolerance OS, extensions,
programmers,applications
Micro-reboots Recovery tolerance OS, applicationsWrappers Isolation / error
detectiontolerance none
Nooks Isolation /error detection/ recovery
tolerance none
isolation, failure detection, and recovery without requiring changes to hardware, extensions, appli-
cations, or programming practices. The OS requires only minor changes for integrating the new
code. Nooks relies on a conventional processor architecture, a conventional programming language,
a conventional operating system architecture, and existing extensions. It is designed to be trans-
parent to the extensions themselves, to support recovery, and to impose only a modest performance
penalty.
28
Sound-Card
Device Driver
OS Kernel
Kernel Interface
Sound Driver
Class Interface
Sound Card
Figure 3.1: A sample device driver.
Chapter 3
DEVICE DRIVER OVERVIEW
A device driver is a kernel-mode software component that provides an interface between the OS
and a hardware device.1 The driver converts requests from the kernel into requests to the hardware.
Drivers rely on two interfaces: the interface that drivers export to the kernel, which provides access
to the device, and the kernel interface that drivers import from the operating system. For example,
Figure 3.1 shows the kernel calling into a sound-card driver to play a tone; in response, the sound
driver converts the request into a sequence of I/O instructions that direct the sound card to emit a
sound.
In addition to processing I/O requests, drivers also handle configuration requests. Configura-
tion requests can change both driver and device behavior for future I/O requests. As examples,
applications may configure the bandwidth of a network card or the volume of a sound card.
In practice, most device drivers are members of a class, which is defined by its interface. Code
that can invoke one driver in the class can invoke any driver in the class. For example, all network
1This thesis uses the terms “device driver” and “driver” interchangeably.
29
drivers obey the same kernel-driver interface, and all sound-card drivers obey the same kernel-driver
interface, so no new kernel or application code is needed to invoke new drivers in these classes. This
class orientation allows the OS and applications to be device-independent, as the details of a specific
device are hidden from view in the driver.
In Linux, there are approximately 20 common classes of drivers. However, not all drivers fit into
classes; a driver may extend the interface for a class with proprietary functions, in effect creating
a new sub-class of drivers. Drivers may also define their own semantics for standard interface
functions, known only to applications written specifically for the driver. In this case, the driver is
in a class by itself. In practice, most common drivers, such as network, sound, and storage drivers,
implement only the standard interfaces.
Device drivers are either request-oriented or connection-oriented. Request-oriented drivers,
such as network drivers and block storage drivers, maintain a single hardware configuration and
process each request independently. In contrast, connection-oriented drivers maintain separate hard-
ware and software configurations for each user of the device. Furthermore, requests on a single
connection may depend on past requests that changed the connection configuration.
We can model device drivers as abstract state machines; each input to the driver from the kernel
or output from the driver reflects a potential state change in the driver. For example, the left side of
Figure 3.2 shows a state machine for a network driver as it sends packets. The driver begins in state
S0, before the driver has been loaded. Once the driver is loaded and initialized, the driver enters
state S1. When the driver receives a request to send packets, it enters state S2, where there is a
packet outstanding. When the driver notifies the kernel that the send is complete, it returns to state
S1. The right side of Figure 3.2 shows a similar state machine for a sound-card driver. This driver
may be opened, configured between multiple states, and closed. The state-machine model aids in
designing and understanding a recovery process that seeks to restore the driver state by clarifying
the state to which the driver is recovering. For example, a mechanism that unloads a driver after a
failure returns the driver to state S0, while one that also reloads the driver returns it to state S1.
30
S2
send complete
S1 S1
S2 S3
open
ioctl-1
ioctl-2
close
close
Network driver Sound-card driver
read,writeread,write
S0 S0
init init
Figure 3.2: Examples of driver state machines.
31
Chapter 4
THE NOOKS RELIABILITY LAYER
4.1 Introduction
Device drivers are a major cause of crashes in commodity operating systems, such as Windows
and Linux, because a fault in a device driver can cause the entire OS to crash. While the core
operating system can be tested broadly, drivers are used by a smaller population and are frequently
updated as new hardware is introduced. Thus, drivers do not achieve the same level of reliability
as core operating system code. Given the vast number of existing device drivers (over 35,000 for
Windows XP), it is not practical to remove bugs from driver code. Instead, the driver reliability
problem demands a structural solution that severs the dependency of operating systems on driver
correctness, so that a driver fault can no longer bring the system down. My work seeks to improve
operating system reliability by allowing the OS to safely execute faulty device drivers. Thus, even
when a driver fails, the system as a whole remains functional.
To this end, I designed and implemented a system called Nooks that isolates device drivers and
then recovers when they fail. Isolation ensures that failed drivers are contained and do not cause
other drivers or the OS itself to fail. Recovery ensures that the system is restored to a functioning
state after a failure. I designed Nooks to be practical, in that supports existing drivers and operating
systems at minimal performance overhead.
Nooks is based on two core design principles:
1. Tolerate with best effort. The system must prevent and recover from most, but not necessarily
all, driver failures.
2. Tolerate mistakes, not abuse. Drivers are generally well behaved but may fail due to errors in
design or implementation.
From the first principle, I am not seeking a complete solution for all possible driver failures. How-
ever, since drivers cause the majority of system failures, tolerating most driver failures will sub-
32
stantially improve system reliability. From the second principle, I have chosen to occupy the design
space somewhere between “unprotected” and “safe.” That is, the driver architecture for conventional
operating systems (such as Linux or Windows) is unprotected: nearly any bug within the driver can
corrupt or crash the rest of the system. In contrast, safe systems (such as SPIN [24] or the Java
Virtual Machine [88]) strictly limit extension behavior and thus make no distinction between buggy
and malicious code. I trust drivers not to be malicious, but I do not trust them not to be buggy.
The practical impact of these principles is substantial, both positively and negatively. On the
positive side, it allows me to define an architecture that directly supports existing driver code with
only moderate performance costs. On the negative side, my solution does not detect or recover from
100% of all possible failures and can be circumvented easily by malicious code acting within the
kernel. As examples, consider a malfunctioning driver that does not corrupt kernel data, but returns
a packet that is one byte short, or a malicious driver that explicitly corrupts the system page table.
Similarly, Nooks cannot safely execute 100% of drivers, because it cannot always distinguish
between correct and faulty behavior. For example, a driver that dereferences and writes to a hard-
coded address containing a kernel data structure may be correct, or it may be accidentally corrupting
the kernel. Nooks cannot tell which is the case. On both dimensions, detecting faulty code and
executing correct code, I strive to maximize Nooks’ abilities while minimizing development and
run-time cost.
Among failures that can crash the system, a spectrum of possible defensive approaches exists.
These range from the Windows approach (i.e., to preemptively crash to avoid data corruption) to
the full virtual machine approach (i.e., to virtualize the entire architecture and provide total isola-
tion). My approach lies in the middle. Like all possible approaches, it reflects trade-offs among
performance, compatibility, complexity, and completeness. Section 4.3.7 describes Nooks’ current
limitations. Some are a result of the current hardware or software implementation while others are
architectural. These limitations have two effects: the inability to tolerate certain kinds of faulty
behavior and the inability to safely execute certain kinds of correct driver code. Despite these lim-
itations, given the tens of thousands of existing drivers, and the millions of failures they cause, a
best-effort solution has great practical value.
At a high level, Nooks is a layer between the operating system kernel and drivers. As shown
in Figure 4.1, the layer is composed of four core services and a controller. In the remainder of this
33
chapter, I discuss the architecture and implementation of the Nooks reliability layer.
4.2 Nooks Architecture
The Nooks architecture seeks to achieve three major goals:
1. Isolation. The architecture must isolate the kernel from driver failures. Consequently, it must
detect these failures before they infect other parts of the kernel.
2. Recovery. The architecture must enable automatic recovery to permit applications that depend
on a failed driver to run after a failure.
3. Backward Compatibility. The architecture must apply to existing systems and existing drivers,
with minimal changes to either.
Achieving all three goals in an existing operating system is challenging. In particular, the need
for backward compatibility rules out certain otherwise appealing technologies, such as type-safe
languages and capability-based hardware. Furthermore, backward compatibility implies that the
performance of a system using Nooks should not be significantly worse than a system without it.
I achieve the preceding goals by creating a new operating system reliability layer that sits be-
tween drivers and the OS kernel. The reliability layer intercepts all interactions between drivers and
the kernel to provide isolation and recovery. A crucial property of this layer is transparency, i.e.,
to be backward compatible, it must be largely invisible to existing components. Furthermore, while
designed for device drivers, it also applies to other kernel extensions, such as network protocols or
file systems.
Figure 4.1 shows this new layer, which I call the Nooks Reliability Layer (NRL). Above the
NRL is the operating system kernel. While the NRL is pictured as a layer, it may require a small
set of modifications to kernel code to be integrated into a particular OS. These modifications need
only be made once. Underneath the NRL is the set of isolated drivers. In general, no modifications
should be required to drivers, since transparency for existing drivers is my major objective.
To support the goals outlined above, the Nooks Reliability Layer provides five major architec-
tural functions: isolation, interposition, object tracking, recovery and control. Isolation ensures that
faulty drivers do not corrupt the kernel. Interposition provides backward compatibility for existing
34
Kernel
Application
Nooks Reliability Layer
Isolation
Application
DeviceDeviceDevice
DriverDriverDriver
ObjectTracking
RecoveryInterposition Control
Figure 4.1: The architecture of Nooks.
drivers by injecting code at the existing kernel-driver interface. Object tracking supports recovery,
by reclaiming those kernel objects in use by a driver after a failure. Recovery ensures that the system
keeps functioning after a driver failure, and control manages the entire layer. I now describe these
functions in more detail.
4.2.1 Isolation
Nooks’ isolation mechanisms prevent driver failures from damaging the kernel (or other isolated
drivers). Every driver in Nooks executes within its own lightweight kernel protection domain. This
domain is an execution context with the same processor privilege as the kernel but with write access
to only a limited portion of the kernel’s address space.
The major task of the isolation mechanism is protection-domain management. This involves the
creation, manipulation, and maintenance of lightweight protection domains. The secondary task is
inter-domain control transfer. Isolation services support the control flow in both directions between
driver domains and the kernel domain.
Unlike system calls, which are always initiated by an application, the kernel frequently calls
into drivers. These calls may generate callbacks into the kernel, which may then generate a call
35
into the driver, and so on. To handle this complex communication, I created a new kernel service,
called Extension Procedure Call (XPC), that is a control transfer mechanism specifically tailored to
isolating extensions within the kernel. This mechanism resembles Lightweight Remote Procedure
Call (LRPC) [23] and Protected Procedure Call (PPC) in capability systems [60]. However, LRPC
and PPC handle control and data transfer between mutually distrustful peers. XPC occurs between
trusted domains but is asymmetric (i.e., the kernel has more rights to the driver’s domain than vice
versa).
4.2.2 Interposition
Nooks’ interposition mechanisms transparently integrate existing drivers into the Nooks environ-
ment. Interposition code ensures that: (1) all driver-to-kernel and kernel-to-driver control flow oc-
curs through the XPC mechanism, and (2) all data transfer between the kernel and driver is managed
by Nooks’ object-tracking code (described below).
The interface between the driver, the reliability layer, and the kernel is provided by a set of
wrapper stubs that are part of the interposition mechanism. Nooks’ stubs provide transparent control
and data transfer between the kernel domain and driver domains. Thus, from the driver’s viewpoint,
the stubs appear to be the kernel’s extension API. From the kernel’s point of view, the stubs appear to
be the driver’s function entry points. Wrappers resemble the stubs in RPC systems [25] that provide
transparent control and data transfer across address space boundaries.
4.2.3 Object Tracking
Nooks’ object-tracking functions oversee all kernel resources used by drivers. In particular, object-
tracking code: (1) maintains a list of kernel data structures a driver may manipulate, (2) controls
all modifications to those structures, and (3) provides object information for cleanup when a driver
fails. Protection domains prevent drivers from directly modifying kernel data structures. Therefore,
object-tracking code must copy kernel objects into a driver domain so they can be modified and then
copy them back after changes have been applied. When possible, object-tracking code verifies the
type and accessibility of each parameter that passes between the driver and kernel. Kernel routines
can then avoid scrutinizing parameters, because wrappers execute checks on behalf of unreliable
36
drivers.
4.2.4 Recovery
Nooks’ recovery functions detect and initiate recovery from a variety of driver failures. Nooks
detects a software failure when a driver invokes a kernel service improperly (e.g., with invalid argu-
ments) or when a driver consumes too many resources. Nooks detects a hardware failure when the
processor raises an exception during driver execution, e.g., when a driver attempts to read unmapped
memory or to write memory outside of its protection domain. A user or a program may also detect
faulty behavior and trigger recovery explicitly.
Unmodified drivers cannot handle their own failure, so Nooks triggers a recovery manager to
restore the system to a functioning state. Recovery managers, which may be specific to a driver,
provide an extensible mechanism for restoring system operation following a failure. Recovery man-
agers are assisted by the isolation and object-tracking services. These services provide functions for
cleanly releasing resources in use by a driver at the time of failure.
4.2.5 Control
Nooks’ control functions are invoked to change the state of a protection domain. Nooks supports
four major control operations: create a protection domain, attach a driver to a protection domain
when the driver is loaded, notify a protection domain of a failure, and destroy a protection domain.
These functions may be accessed externally, through tools such as the system module loader, or
internally, when invoked by Nooks to recover after a failure.
4.2.6 Summary
The Nooks Reliability Layer improves operating system reliability by isolating and recovering from
the failure of existing, unmodified device drivers. The layer is composed of a controller and four
core services: isolation, interposition, object tracking, and recovery. These services prevent device
drivers from corrupting the kernel, detect driver failures, and restore system functionality after a
failure.
37
4.3 Implementation
I implemented Nooks within the Linux 2.4.18 kernel on the Intel x86 architecture. I chose Linux
as a platform because of its popularity and its wide support for kernel extensions in the form of
loadable modules. Linux proved to be a challenging platform, because the kernel provides over
700 functions callable by extensions and more than 650 extension-entry functions callable by the
kernel. Moreover, few data types are abstracted, and drivers directly access fields in many kernel
data structures. Despite these challenges, I brought the system from concept to function in about 18
months. The implementation supports both device drivers and other kernel extensions.
The Linux kernel supports standard interfaces for many extension classes. For example, there is
a generic interface for block, character, and network device drivers, and another one for file systems.
The interfaces are implemented as C language structures containing a set of function pointers.
Most interactions between the kernel and drivers take place through function calls, either from
the kernel into drivers or from drivers into exported kernel routines. Drivers directly access several
global data structures, such as the current task structure and the jiffies time counter. Fortunately,
they modify few of these structures, and frequently do so through preprocessor macros and inline
functions. As a result, Nooks can interpose on most kernel-driver interactions by intercepting the
function calls between the driver and kernel.
Figure 4.2 shows the Nooks layer in Linux. Under the Nooks controller and object tracker are
isolated kernel extensions: a single device driver, three cooperating drivers, and a kernel service.
These extensions are wrapped by Nooks’ wrapper stubs, as indicated by the shaded boxes sur-
rounding them. Each wrapped box, containing one or more extensions, represents a single Nooks
protection domain. Shown at the top, in user mode, is a recovery manager, which is responsible for
restarting extensions after a failure. Figure 4.2 also shows unwrapped drivers and kernel services
that continue to interface directly to the Linux kernel.
To facilitate portability, Nooks does not use the Intel x86 protection rings or memory segmenta-
tion mechanisms. Instead, drivers execute at the same privilege level as the rest of the kernel (ring
0). A conventional page table architecture provides memory protection, which can be implemented
both with hardware- and software-filled TLBs.
Table 4.1 shows the size of the Nooks implementation. As discussed in Chapter 6, this code can
38
Applications Daemons
Linux Kernel
Nooks Controller
Wrappers
Driver
Driver
Driver
Wrappers
Wrappers
Driver
KernelService
KernelService
KernelService
KernelService
Driver
Device Device Device
ApplicationsApplications
DaemonsDaemonsRestart
Recovery Manager
Object Tracker
Figure 4.2: The Nooks Reliabilty Layer inside the Linux OS.
Table 4.1: The number of non-comment lines of source code in Nooks.
Source Components # LinesDomain Management 2,391Object Tracking 1,498Extension Procedure Call 928Wrappers 14,484Recovery 1,849Build tools 1,762Linux Kernel Changes 924Miscellaneous 1,629
Total number of lines of code 25,465
tolerate the failure of fourteen device drivers and two other kernel extensions. The Nooks reliability
layer comprises less than 26,000 lines of code. In contrast, the kernel itself has 2.4 million lines, and
the Linux 2.4 distribution has about 30 million [206]. Other commodity systems are of similar size.
For example, Microsoft Windows Server 2003 operating system contains over 50 million lines of
code [193]. Clearly, relative to a base kernel, its drivers, and other extensions, the Nooks reliability
layer introduces only a modest amount of additional system complexity.
In the following subsections I discuss the implementation of Nooks’ major components: isola-
39
Linux Kernel
Driver 1
Driver 1: RDriver 2: R/W
Kernel: R/W
Driver 2
Heap Stacks
I/O
Buffers
Object
Table
Heap Stacks
I/O
Buffers
Object
Table
Driver 1: R/WDriver 2: R
Kernel: R/W
Driver 1: RDriver 2: R
Kernel: R/W
Figure 4.3: Protection of the kernel address space.
tion, interposition, wrappers, object tracking, and recovery. I describe wrappers separately because
they comprise the bulk of Nooks’ code and complexity. Finally, I describe limitations of the Nooks
implementation.
4.3.1 Isolation
The isolation component of Nooks consists of two parts: (1) memory management to implement
lightweight kernel protection domains with virtual memory protection, and (2) Extension Procedure
Call (XPC) to transfer control safely between extensions and the kernel.
Figure 4.3 shows the Linux kernel with two lightweight kernel protection domains, each con-
taining a single driver. All components exist in the kernel’s address space. However, memory access
rights differ for each component: e.g., the kernel has read-write access to the entire address space,
while each driver is restricted to read-only kernel access and read-write access to its local domain.
This is similar to the management of address space in single-address-space operating systems, where
all applications share the virtual-to-physical mapping of addresses but differ in their access rights to
memory [40].
To provide drivers with read access to the kernel, Nooks’ memory management code maintains
a synchronized copy of the kernel page table for each domain. Each lightweight protection domain
has private structures, including a domain-local heap, a pool of stacks for use by the driver, memory-
mapped physical I/O regions, and kernel memory buffers, such as socket buffers or I/O blocks, that
are currently in use by the driver.
40
Nooks uses the extension procedure call (XPC) mechanism to transfer control between driver
and kernel domains. The wrapper mechanism, described in Section 4.3.3, makes the XPC mech-
anism invisible to both the kernel and drivers, which continue to interact through their original
procedural interfaces.
Two functions internal to Nooks manage XPC control transfer: (1) nooks driver call
transfers from the kernel into a driver, and (2) nooks kernel call transfers from drivers into
the kernel. These functions take a function pointer, an argument list, and a protection domain. They
execute the function with its arguments in the specified domain. The transfer routines save the
caller’s context on the stack, find a stack for the calling domain (which may be newly allocated or
reused when calls are nested), change page tables to the target domain, and then call the function.
XPC performs the reverse operations when the call returns.
Changing protection domains requires a change of page tables. The Intel x86 architecture flushes
the TLB on such a change, hence, there is a substantial cost to entering a lightweight protection
domain, both from the flush and from subsequent TLB misses. This cost could be mitigated in a
processor architecture with a tagged TLB, such as the MIPS or Alpha, or with single-address-space
protection support [123], such as the IA-64 or PA-RISC. However, because Nooks’ lightweight
protection domains execute on kernel threads that share the kernel address space, they reduce the
costs of scheduling and data copying on a domain change when compared to normal cross-address
space or kernel-user RPCs.
To reduce the performance cost of XPC, Nooks supports deferred calls, which batch many calls
into a single domain crossing. Nooks can defer function calls that have no visible side effects to the
call. Wrappers queue deferred function calls for later execution, either at the entry or exit of a future
XPC. Each domain maintains two queues: a driver-domain queue holds delayed kernel calls, and
a kernel-domain queue holds delayed driver calls. As an example, I changed the packet-delivery
routine used by the network driver to batch the transfer of message packets from the driver to the
kernel. When a packet arrives, the driver calls a wrapper to pass the packet to the kernel. The
wrapper queues a deferred XPC to deliver the packet after the driver completes interrupt processing.
In addition to deferring calls for performance reasons, Nooks also uses deferred XPC to synchro-
nize some objects shared by the kernel and drivers. In Linux, the kernel often provides a pointer that
the driver may modify with no explicit synchronization. The Linux kernel is non-preemptive and
41
assumes, safely, that the modification is atomic and that the driver will update it “in time.” In such
cases, the wrapper queues a deferred function call to copy the modified object back when the driver
returns control to the kernel.
I previously noted that Nooks protects against bugs but not against malicious code. Lightweight
protection domains reflect this design. For example, Nooks prevents a driver from writing kernel
memory, but it does not prevent a malicious driver from replacing the domain-local page table
explicitly by reloading the hardware page table base register.
Furthermore, Nooks currently does not protect the kernel from DMA by a device into the kernel
address space. Preventing a rogue DMA requires hardware that is not generally present on x86 com-
puters. However, Nooks tracks the set of pages writable by a driver and could use this information
to restrict DMA on a machine with the suitable hardware support.
I made several one-time changes to the Linux kernel to support isolation. First, to maintain
coherency between the kernel and driver page tables, I inserted code wherever the Linux kernel
modifies the kernel page table. Second, I modified the kernel exception handlers to detect exceptions
that occur within Nooks’ protection domains. This new code swaps in the kernel’s stack pointer and
page table pointer for the task. On returning from an exception, the code restores the stack pointer
and page table for the driver. Finally, because Linux co-locates the task structure on the kernel
stack (which XPC replaces with a driver stack), I had to change the mechanism for locating the
current task structure. I currently use a global variable to hold the task pointer, which is sufficient
for uniprocessor systems. On a multiprocessor, I would follow the Windows practice of storing the
address in an otherwise unused x86 segment register.
4.3.2 Interposition
Interposition allows Nooks to intercept and control communication between drivers and the kernel.
Nooks interposes on kernel-driver control transfers with wrapper stubs. Wrappers provide trans-
parency by preserving existing kernel-driver procedure-call interfaces while enabling the protection
of all control and data transfers in both directions. Control interposition required two changes to
Linux kernel code. First, I modified the standard module loader to bind drivers to wrappers instead
of kernel functions when loading a driver. As a side benefit, the module loader need not bind drivers
42
Kernel
Driver 3Driver 2Driver 1
write()
write()Wrapper
Kernel
TrampolineTrampolineTrampoline
write()
Driver 3Driver 2Driver 1
write()
Figure 4.4: Trampoline wrappers for invoking multiple implementations of an interface function.
against wrappers. Thus, an administrator may choose whether or not to isolate each driver. This
feature allows Nooks to isolate failure-prone drivers while leaving reliable ones unaffected.
Second, I modified the kernel’s module initialization code to explicitly invoke Nooks on the ini-
tialization call into a driver, enabling the driver to execute within its lightweight protection domain.
Following initialization, wrappers replace all function pointers passed from the driver to the kernel
with pointers to wrappers instead. This causes the kernel to call wrapper functions instead of driver
functions directly.
Nooks cannot interpose on all functions exported by drivers with a single wrapper stub be-
cause the kernel selects the driver it is invoking by the address of function it calls. The left side of
Figure 4.4 shows the kernel selecting among multiple instances of a driver function based on the
function address. Thus, a unique function pointer is required for each implementation of a function
in the kernel-driver interface. As there is only one implementation of kernel functions, a single
wrapper for each function suffices. Functions exported by drivers, though, may be implemented
multiple times in different drivers or even within the same driver. As a result, a separate function
pointer is required for each instance of a function in the driver interface.
To address this problem, Nooks generates “trampoline” code at run-time that pushes a pointer
43
Kernel
Driver
Wrapper
XPC
Copy / Sync.
Check
Driver
Kernel
XPC
Wrapper
Check
Copy / Sync.
Figure 4.5: Control flow of driver and kernel wrappers.
to the driver instance on the stack and then invokes a wrapper that is common to all instances of
the function. Nooks passes the kernel a pointer to the trampoline instead of the common wrapper
function. The right side of Figure 4.4 shows the kernel invoking trampolines, which pass control to
a common wrapper function. This mechanism ensures that all instances of a driver function have a
unique function pointer and that the wrapper logic is centralized in a single implementation.
In addition to interposing on control transfers, Nooks must interpose on some data references.
Drivers are linked directly to global kernel variables that they read but do not write (e.g., the current
time). For global variables that drivers modify, I create a shadow copy of the kernel data structure
within the driver’s domain that is synchronized to the kernel’s version. For example, Nooks uses
this technique for the softnet data structure, which contains a queue of the packets sent and
received by a network driver. The object tracker synchronizes the contents of the kernel and driver
version of this structure before and after XPCs into a network driver.
4.3.3 Wrappers
As noted above, Nooks inserts wrapper stubs between kernel and driver functions. There are two
types of wrappers: kernel wrappers, which are called by drivers to execute kernel-supplied func-
44
tions; driver wrappers, which are called by the kernel to execute driver-supplied functions. In both
cases, a wrapper functions as an XPC stub that appears to the caller as if it were the target procedure
in the called domain.
Both wrapper types perform the body of their work within the kernel’s protection domain. There-
fore, the domain change occurs at a different point depending on the direction of transfer, as shown
in Figure 4.5. When a driver calls a kernel wrapper, the wrapper performs an XPC on entry so that
the body of the wrapper (i.e., object checking, copying, etc.) can execute in the kernel’s domain.
Once the wrapper’s work is done, it calls the target kernel function directly with a regular procedure
call. In the opposite direction, when the kernel calls a driver wrapper, the wrapper executes within
the kernel’s domain. When it is done, the wrapper performs an XPC to transfer to the target function
within the driver.
Wrappers perform three basic tasks. First, they check parameters for validity by verifying with
the object tracker and memory manager that pointers are valid. Second, they implement call-by-
value-result semantics for XPC, by creating a copy of kernel objects on the local heap or stack
within the driver’s protection domain. These semantics ensure that updates to kernel objects are
transactional, because they are only applied after the driver completes, when the wrappers copy the
results back to the kernel. Third, wrappers perform an XPC into the kernel or driver to execute the
desired function, as shown in Figure 4.5.
While wrappers must copy data between protection domains, no marshaling or unmarshaling
is necessary, because the driver and kernel share the same address space. Instead, wrappers may
directly allocate and reference memory in either the kernel or the driver protection domains. The
code for synchronizing simple objects is placed directly in the wrappers, while the object tracker
provides synchronization routines for complex objects with many pointers. As an optimization,
wrappers may pass parameters that are only read but not written by drivers without modification, as
any attempt to modify the parameter will cause a memory access fault.
To improve performance, the wrappers rely on several techniques for moving complex objects
between protection domains. In some cases, Nooks copies objects into the driver’s protection do-
main, following embedded pointers as appropriate. From an analysis of driver and extension source
code, I determined that it is generally unnecessary to copy the complete transitive closure of an
object; while drivers read pointers more than one level removed from a parameter, they generally
45
do not write to them. In other cases, Nooks avoids copying entirely by changing the protection on
the page containing an object. A “page tracker” mechanism within the object tracker remembers
the state of these mapped pages and grants and revokes driver access to the pages. Nooks uses this
mechanism to avoid copying network packets and disk blocks.
Opaque pointers, which appear as void * in type declarations, pose a challenge for isolation.
The contents of these pointers cannot be copied between protection domains because their type is
not known. In practice, though, opaque pointers can often be interpreted based on how they are
used. There are three classes of opaque pointers:
1. Private to driver. These pointers, commonly used as parameters to callback functions, are
never dereferenced by any code other than the driver.
2. Shared with applications. These pointers are dereferenced by user-mode applications, and
hence must point into user address space.
3. Shared with other drivers. These pointers are opaque to the kernel but may be dereferenced
by other drivers.
Any pointer dereferenced by the kernel is not truly opaque, because Nooks can reuse the kernel’s
logic to determine its type.
The first class of opaque pointers, private to driver, may be passed through wrappers and the
object tracker without any work because the kernel never dereferences the pointer. Pointers in the
second class, shared with applications, reference objects from user space. Hence, drivers cannot di-
rectly dereference these pointers and must instead use accessor functions, such as copy from user.
Nooks interposes on these function to grant drivers access to user-mode memory.
The third class of opaque pointers, shared with other drivers, poses a problem because drivers
may attempt to modify the objects to which they point. As a result, drivers using this type of opaque
pointer cannot be isolated from each other; they must all execute in the same protection domain.
Wrappers are relatively straightforward to write and integrate into the kernel. I developed a tool
that automatically generates trampoline code and the skeleton of wrapper bodies from Linux kernel
header files. To create the wrappers for exported kernel functions, the tool takes a list of kernel
function names and generates wrappers that implement function interposition and XPC invocations.
46
Wrappers Used By Extensions
0
20
40
60
80
100
120
3c59
x
3c90
x
e100
e100
0
epic10
0
pcne
t32
audigy
cs42
32
emu1
0k1
es13
71
i810
_aud
io sb
ide-
disk
ide-
cdvfat
khttpd
Extension
Nu
mb
er o
f W
rap
pers U
sed
Common to all
extensions
Common to this
driver class
Shared with at least
one extension
Unique to this
extension
Network Drivers Sound Drivers IDE
Drivers
Other
Figure 4.6: Code sharing among wrappers for different extensions.
Similarly, for the kernel-to-driver interface, the tool takes a list of interfaces (C structures containing
function pointers) and generates wrappers for the kernel to call.
I wrote the body of wrapper functions manually. This is a one-time task required to support
the kernel-driver interface for a specific OS. Once written, wrappers are automatically usable by
all drivers that use the kernel’s interface. This code checks parameters for simple errors, such as
pointing to invalid addresses, and then moves parameters between protection domains.
Writing a wrapper requires knowing how drivers use a parameter: whether it is live across
multiple calls to the drivers, whether it can be passed to other threads or back to the kernel, and
which fields of the parameter can be modified. While I performed this analysis manually, static
analysis tools could determine these properties by analyzing an existing set of drivers [69].
47
Wrapper Code Sharing
Previously, Table 4.1 showed that the Nooks implementation includes 14K lines of wrapper code,
over half of the Nooks code base. Thus, it is important to understand the effort required to write
these wrappers when isolating additional drivers or kernel extensions.
Chapter 6 describes in more detail the sixteen extensions I isolated: six sound-card drivers,
six Ethernet drivers, two IDE storage drivers, a file system (VFAT), and an in-kernel Web server
(kHTTPd). I only implemented wrappers for functions that were actually used by these extensions.
Thus, I wrote wrappers for 126 of the 650 functions imported by the kernel and for 329 of the 700
exported functions, totaling 455 wrappers.
Figure 4.6 shows the total number of wrappers (both kernel and driver wrappers) used by each
of these extensions. Each bar gives a breakdown of the number of wrappers unique to that extension
and the number of wrappers shared in various ways. Sharing reduces the cost of isolating a given
extension. For example, of the 46 wrappers used by the pcnet32 Ethernet driver (35 kernel wrappers
and 11 extension wrappers), 22 are shared among the six network drivers. Similarly, 31 wrappers
are shared between the six sound-card drivers. Overall, of the 158 wrappers that are not shared,
116 are in the VFAT and kHTTPd extensions. While VFAT would share many of its wrappers with
other file systems, kHTTPd is a one-of-a-kind extension and as a result its wrappers are unlikely to
be shared. Thus, the class structure of drivers reduces the effort required to isolate a driver. Once
Nooks supports a single driver in a class, little work is needed to support additional drivers.
4.3.4 Object Tracking
The object tracker facilitates the recovery of kernel objects following a driver failure. The Nooks
object tracker performs two independent tasks. First, it records the addresses of all objects in use by
a driver in a database. As an optimization, objects used only for the duration of a single XPC call are
recorded on the kernel stack. Objects with long lifetimes are recorded in a per-protection-domain
hash table. Second, for objects that drivers may modify, the object tracker creates and manages
a driver version of the object and records an association between the kernel and driver versions.
Wrappers rely on this association to map parameters between the driver’s protection domain and the
kernel’s protection domain.
48
The Nooks implementation currently supports many kernel object types, such as tasklets, PCI
devices, inodes, and memory pages. To determine the set of objects to track, I inspected the inter-
faces between the kernel and the supported drivers and noted every object type that passed through
those interfaces. I then wrote object-tracking procedures for each of the 52 object types that I found.
For each object type, there is a unique type identifier and code to release instances of that type
during recovery. Complex types also have a routine to copy changes between a kernel and driver
instance of the type.
When an object “dies” and is no longer usable by a driver, the object tracker must remove the
object from its database. Determining when an object will no longer be used requires a careful
examination of the kernel-driver interface. This task is possible because the kernel requires the
same information to safely reclaim shared objects. For example, some objects are accessible to the
driver only during the lifetime of a single XPC call from the kernel. In this case, I add the object
to the tracker’s database when the call begins and remove it on return. Other objects are explicitly
allocated and deallocated by the driver, in which case I know their lifetimes exactly. In still other
cases, I go by the semantics of the object and its use. For example, drivers allocate the timer data
structure to register for a future callback. I add this object to the object tracker when a driver calls
add timer and remove it when the timer fires, at which point I know that it is no longer used.
The object-tracking code is conservative, in that it may under-estimate the lifetime of an object
and unnecessarily add and remove the same object from the database multiple times. It will not,
however, allow a driver to access an object that the kernel has released.
In addition to tracking objects in use by drivers, the tracker must record the status of locks that
are shared with the kernel. When a driver fails, Nooks releases all locks acquired by the driver to
prevent the system from hanging. As a result, calls to lock kernel data structures require an XPC into
the kernel to acquire the lock, synchronize the kernel and driver versions of the data structure, and
record that the lock was acquired. Fortunately, the performance impact of this approach is minimal
because shared locks are uncommon in the Linux device drivers I examined.
49
4.3.5 Recovery
The recovery code in Nooks consists of three components. First, the isolation components detect
driver failures and notify the controller. Second, the object tracker and protection domains support
cleanup operations that release the resources in use by a driver. This functionality is available to the
third component, a recovery manager, whose job is to recover after a failure. The recovery manager
may be customized to a specific driver or class of drivers, as I discuss more in Chapter 5.
Fault Detection
Nooks triggers recovery when it detects a failure through software checks (e.g., parameter validation
or livelock detection), processor exceptions, or notification from an external source. Specifically,
the wrappers, protection domains, and object tracker notify the controller of a failure when:
• The driver passes a bad parameter to the kernel, such as accessing a resource it had freed or
unlocking a lock not held.
• The driver allocates too much memory, such as an amount exceeding the physical memory in
the computer.
• The driver executes too frequently without an intervening clock interrupt (implying livelock).
• The driver generates an invalid processor exception, such as an illegal memory access or an
invalid instruction.
In addition, it is possible to implement an external failure detector, such as a user- or kernel-mode
agent, that notifies the controller of a failure. In all cases, the controller invokes the driver’s recovery
manager.
Recovery Managers
The recovery manager is tasked with returning the system to a functioning state. Nooks initially
supported two simple recovery managers. The default recovery manager is a kernel service that
simply unloads the failed driver, leaving the system running but without the services of the driver.
The restart recovery manager is a user-mode agent that similarly unloads the failed driver but then
executes a script to reload and restart the driver. In the next chapter, I discuss a third recovery
50
manager that provides additional services. For each protection domain, the Nooks controller records
which manager to notify when a failure occurs.
The XPC, object tracking, and protection domain code all provide interfaces to the recovery
managers. The XPC service allows a manager to signal all the threads that are currently executing
in the driver or have called through the driver and back into the kernel. The signal causes the threads
to unwind out of the driver by returning to the point where they invoked the driver without executing
any additional driver code. Threads that are in a non-interruptible state may ignore the signal, and
hence complete recovery is impossible if a sleeping thread never wakes. It is not safe to wake such
a thread, because the condition it is waiting for may not be true, and the thread could then corrupt
data outside the driver’s protection domain. By ignoring hung threads, partial recovery may still be
possible. Uninterruptible sleeps are infrequent in the Linux kernel, however, so I do not believe this
to be a significant limitation.
The object tracker provides an interface to recovery managers to enumerate the objects in use by
a driver at the time of failure and to garbage collect the objects by releasing them to the kernel. The
manager may choose both the set of objects it releases and the order in which to release them. Thus,
it may preserve objects for use by the driver after recovery, such as memory-mapped I/O buffers that
a hardware device continues to access.
Lightweight kernel protection domains provide similar support for recovery. The domains record
the memory regions accessible to a driver and provide interfaces for enumerating the regions and
for releasing the regions to the kernel.
Sequence of Recovery Actions
The Nooks recovery managers execute in stages to first disable a failed driver, then unload the driver,
and optionally restart the driver. These actions are derived from the state machine model of device
drivers in Chapter 3. The first two stages return the system to the initial state, S0, where the driver is
not present. The final stage, restarting the driver, returns the driver to state S1, where it is initialized
and ready to process requests.
Once a failure has been detected, the first step of recovery is to disable the failed driver. As
Nooks isolation is imperfect, this step ensures that the driver does not cause any further corruption.
51
At this stage, the hardware device must be disabled to prevent it from corrupting system state as
well.
The second stage of recovery is to unload the failed driver and return the system (or rather,
the driver-related portions of the system) to its state before the driver was loaded. The recovery
manager invokes the object tracker to release objects in use by the driver, protection domain code to
release memory objects, and the XPC code to signal threads that were executing in the driver. The
default recovery manager completes execution after this stage, leaving the OS running but without
the services of the failed driver.
To further improve reliability, the restart recovery manager continues on to reload and re-
initialize the failed driver. While applications that were actively using the old driver may experience
failures when the driver is unloaded, applications not using the driver continue to execute properly
and those launched after the failure may use the driver.
Complications
While conceptually simple, the practicalities of dealing with low-level device driver code pose many
complications to the recovery process. The recovery manager must ensure that the kernel is returned
to a functioning state and that devices managed by a failed driver do not interfere with recovery.
As much as possible, the recovery code attempts to mimic the behavior a driver that is shutting
down cleanly. Hence, the object tracker uses the same routines that drivers use to release objects.
For example, the object tracker calls the kfree skb function to release network packets in flight,
which decrements a reference count and potentially frees the packet. As a result, higher levels of
software, such as network protocol code, continue to function correctly during and after recovery.
The recovery manager must ensure that the system is not left in a hung state after recovery. Two
factors prevent the system from hanging when an extension fails while holding a lock. First, if the
lock is shared with the kernel, the object tracker records the lock status so that a recovery manager
can release the lock. Second, if the lock is private to the driver, any threads blocked on the lock will
be signaled to unwind by the XPC service.
As part of the unloading process, the recovery manager must ensure that open connections to
drivers and other extensions are invalidated, because the connection state is lost when the failed
52
driver is unloaded. However, the driver’s clients may still attempt to use these connections. Thus,
the object tracker, when releasing connection objects, replaces the function table in the connection
with a table of dummy functions. These functions return errors for all calls other than the release
operation, which closes the connection. As a result, clients can close a connection to a failed driver
but will receive errors if they try to request any other services on a connection.
In the period between the detection of a failure and the unloading of the driver, requests may
continue to arrive. To ensure that driver code is not invoked after a failure, wrappers check the state
of a driver before performing an XPC into a driver. If the driver has failed, the wrapper returns an
error code to its caller. After the driver is unloaded, attempts to access the driver will receive an
error from the kernel, because the driver no longer exists.
Device hardware may continue to access system resources even after its driver has been halted.
For example, a device may continue to generate interrupts or DMA into system. Thus, the recovery
manager disables interrupt processing for the device controlled by the driver during the first stage of
recovery. This prevents livelock that could occur if device interrupts are not properly dismissed. The
recovery manager must also delay releasing resources, such as physical memory, that a device may
access until its driver has recovered and is in control of the device. For example, a network device
may continue to write to packets in memory buffers; therefore, those buffers cannot be released until
the driver has restarted and reinitialized the device. The default recovery manager, because it does
not reload drivers, cannot reclaim these resources.
The recovery code must also take care to prevent denial-of-service when multiple drivers share
a single interrupt request line. Disabling an interrupt blocks all drivers that share the line. Nooks
addresses this problem with a timer that polls shared interrupts during recovery. Under normal con-
ditions, all drivers using the shared interrupt are invoked every time any device raises the interrupt
line. Each driver checks whether its device raised the interrupt, and if so, handles it appropriately.
Otherwise, the driver ignores the interrupt. Nooks relies on this behavior during recovery by invok-
ing all drivers sharing an interrupt when the timer fires. If a device did generate an interrupt during
the timer period, its driver will handle it. Otherwise, it will ignore the call. While performance is
degraded during recovery because of the extra latency induced by the timer, this approach prevents
livelock due to blocking interrupts.
53
4.3.6 Achieving Transparency
As mentioned previously, a critical goal for Nooks is transparency, meaning that Nooks must be able
to isolate existing drivers written with no knowledge of Nooks. This allows drivers to run unchanged
and yet still benefit from the system’s fault isolation and recovery mechanisms.
Two components of Nooks’ facilitate transparency for existing drivers:
1. Nooks provides wrapper stubs for every function call in the kernel-driver interface.
2. Nooks provides object-tracking code for every object type that passes between the driver and
the kernel.
The wrapper stubs interpose on every call between a driver and the kernel and invoke object-tracking
code to manage every parameter passed between the caller and callee. The wrappers also invoke the
XPC mechanism to transfer control from the caller’s domain to the callee’s domain. Thus, wrapper
stubs allow isolation code to be interposed transparently between drivers and the kernel. The object-
tracking code allows drivers to access kernel objects as expected without explicit code in the driver
to copy data between protection domains. Thus, driver code and the kernel code invoking the driver
need not be changed.
Neither the driver nor the kernel is aware of the existence of the Nooks layer. For the fourteen
device drivers I isolated with Nooks, none required source code changes. Of the two other kernel
extensions, only kHTTPd, the kernel-mode web server, required changes, and then only to 13 lines
of code. With these small changes, Nooks can isolate the extension and track all its resources,
allowing Nooks to catch errant behavior and to clean up during recovery. Furthermore, of the 924
lines of kernel changes mentioned in Table 4.1, none were to code that invokes drivers; they were
primarily changes to memory management and interrupt handling code.
While Nooks does not require changes to driver source code, it does require that drivers be
recompiled to convert inline functions and macros into XPC invocations. Because Linux does not
support a binary interface to the kernel (extensions must be recompiled for every kernel), this is not
a major limitation.
54
4.3.7 Implementation Limitations
Section 4.1 described the Nooks philosophy of designing for mistakes and for best effort. The Nooks
implementation involves many trade-offs. As such, it does not provide complete isolation or fault
tolerance for all possible driver failures. As previously discussed, these limitations are either faulty
behavior that cannot be contained or correct behavior that cannot be recognized as such.
The faulty behavior that Nooks cannot detect stems from two factors. First, Nooks runs drivers
in kernel mode to simplify backward compatibility, so Nooks cannot prevent drivers from deliber-
ately executing privileged instructions that corrupt system state. In addition, Nooks cannot prevent
infinite loops within the driver when interrupts are disabled, because it cannot regain control of the
processor from the driver. Second, Nooks cannot fully verify the behavior of drivers and other ker-
nel extensions, so Nooks’ parameter checks are incomplete. For example, Nooks cannot determine
that a network driver has actually sent the packet when it notifies the kernel that it has done so.
Other limitations stem from driver and extension behavior that Nooks cannot execute safely.
Nooks, as an architecture, is limited to extension mechanisms with well-typed parameters. Nooks
must be able to determine the type of any data passed out by the driver to the kernel or another driver
in order to completely interpose on the driver’s communication. While this is often the case, if it is
not, Nooks cannot ensure that an isolated driver does not pass corrupt data to another driver.
Nooks’ restart recovery manager restores system functionality by restarting the failed driver. As
a result, it can only recover for drivers that can be killed and restarted safely. This is true for device
drivers that are loaded and unloaded by the kernel’s hotplug mechanism, but may not be true for all
drivers or kernel extensions.
Finally, Nooks interposes on the kernel-driver interface when drivers are loaded into the kernel.
Integrated drivers that are compiled into the kernel, such as real-time clock timers and interrupt
controllers, require additional changes to the kernel to be initialized into a protection domain. In
addition, these drivers’ global variables must be placed on pages separate from kernel globals so the
driver is granted write access only to its globals but not others that lie nearby in memory. While
these changes increase the amount of kernel code that must change to support Nooks, they are not
fundamental limitations of the Nooks architecture.
These limitations are not insignificant, and crashes may still occur. Nonetheless, the implemen-
55
tation will allow a kernel to resist many crashes caused by drivers. Given the enormous number of
such crashes, a best-effort solution can have a large impact on overall reliability.
4.4 Summary
Device drivers are a major source of failure in modern operating systems. Nooks is a new reliability
layer intended to significantly reduce driver-related failures. Nooks isolates drivers in lightweight
kernel protection domains and relies on hardware and software checks to detect failures. After
a failure, Nooks recovers by unloading and then reloading the failed driver. Nooks focuses on
achieving backward compatibility, that is, it sacrifices complete isolation and fault tolerance for
compatibility and transparency with existing kernels and drivers. As a result, Nooks has the potential
to greatly improve the reliability of today’s operating systems by removing their dependence on
driver correctness.
This chapter presented the details of the Nooks isolation mechanisms. Nooks is composed of a
controller and four core services: isolation, interposition, object tracking, recovery. These services
allow Nooks to transparently isolate existing drivers from the kernel, detect driver failures, and
recover after a driver failure. The effectiveness of Nooks depends on its abilities to isolate and
recover from driver failures, which I evaluate in Chapter 6.
56
Chapter 5
THE SHADOW DRIVER RECOVERY MECHANISM
5.1 Introduction
The Nooks reliability layer recovers from driver failure by unloading and then reloading the failed
driver. As I will show in Section 6.3, reloading failed drivers is effective at preventing system
crashes. However, users of a computer are not solely interested in whether the operating system
continues to function. Often, users care more about the applications with which they interact. If
applications using drivers fail, then I have only partially achieved my goal of improving reliability.
With the restart recovery manager, calls into a driver that fails and subsequently recovers may
return error codes because the recovery manager unloads the driver and invalidates open connections
to the driver during recovery. As a result, clients of a recovered driver would themselves fail if they
depend on the driver during or after recovery. For example, audio players stopped producing sound
when a sound-card driver failed and recovered. For the same reason, Nooks cannot restart drivers
needed by the kernel, such as disk drivers. Requests to the disk driver fail while the driver is
recovering. When the Linux kernel receives multiple errors from a disk driver used for swapping, it
assumes that the device is faulty and crashes the system.
In addition, any settings an application or the OS had downloaded into a driver are lost when the
driver restarts. Thus, even if the application reconnects to the driver, the driver may not be able to
process requests correctly.
These weaknesses highlight a fundamental problem with a recovery strategy that reveals driver
failures to their clients: the clients may not be prepared to handle these failures. Rather, they are
designed for the more common case that either drivers never fail, or, if they fail, the whole system
fails.
To address these problems, I developed a new recovery mechanism, called a shadow driver, or
more simply, a shadow. My design for shadows reflects four principles:
57
1. Device driver failures should be concealed from the driver’s clients. If the operating sys-
tem and applications using a driver cannot detect that it has failed, they are unlikely to fail
themselves.
2. Driver recovery logic should be generic. Given the huge number and variety of device drivers,
it is not practical to implement per-driver recovery code. Therefore, the architecture must
enable a single shadow driver to handle recovery for a large number of device drivers.
3. Recovery logic should be centralized. The implementation of the recovery system should be
simplified by consolidating recovery knowledge within a small number of components.
4. Recovery services should have low overhead when not needed. The recovery system should
impose relatively little overhead for the common case (that is, when drivers are operating
normally).
Overall, these design principles aim to protect applications and the OS from driver failure, while
minimizing the cost required to make and use shadow drivers.
Shadow drivers do not apply to all types of kernel extensions or failures. First, shadow drivers
only apply to device drivers that belong to a class and share a common calling interface. Second,
shadow drivers recover after a failure by restarting the driver and replaying past requests and hence,
can only recover from failures that are both transient and fail-stop. Deterministic failures may recur
when the driver recovers, again causing a failure. Recoverable failures must be fail-stop, because
shadow drivers must detect a failure in order to conceal it from the OS and applications. Hence,
shadow drivers require an isolation subsystem to detect and stop failures before they are visible to
applications or the operating system.
In the remainder of this chapter, I describe the architecture and implementation of the shadow
driver recovery mechanism.
5.2 Shadow Driver Design
Shadow drivers are a new recovery architecture that takes advantage of the shared properties of a
class of drivers for recovery. The architecture consists of three components: shadow drivers, taps,
58
and a shadow recovery manager. I now describe these components in more detail.
5.2.1 Shadow Drivers
A shadow driver is a piece of software that facilitates recovery for an entire class of device drivers.
A shadow driver instance is a running shadow driver that recovers for a single, specific driver. The
shadow instance compensates for and recovers from a driver that has failed. When a driver fails, its
shadow restores the driver to its pre-failure state. This allows, for example, the recovered driver to
complete requests made before the failure.
Shadow drivers rely on the state machine model of drivers discussed in Chapter 3. Whereas the
previous Nooks recovery managers seek to restore the driver to its unloaded state or initialized state,
shadow drivers seek to restore drivers to their state at the time of failure.
Shadow drivers execute in one of two modes: passive or active. Passive mode is used during
normal (non-faulting) operation, when the shadow driver monitors all communication between the
kernel and the device driver it shadows. This monitoring is achieved via replicated procedure calls:
a kernel call to a device driver function causes an automatic, identical call to the corresponding
shadow driver function. Similarly, a driver call to a kernel function causes an automatic, identical
call to a corresponding shadow driver function. These passive-mode calls are transparent to the
device driver and the kernel and occur only to track the state of the driver as necessary for recovery.
Based on the calls, the shadow tracks the state transitions of the shadowed device driver.
Active mode is used during recovery from a failure. Here, the shadow performs two functions.
First, it “impersonates” the failed driver, intercepting and responding to calls for service. Therefore,
the kernel and higher-level applications continue operating as though the driver had not failed. Sec-
ond, the shadow driver restarts the failed driver and brings it back to its pre-failure state. While the
driver restarts, the shadow impersonates the kernel to the driver, responding to its requests for ser-
vice. Together, these two functions hide recovery from the driver, which is unaware that a shadow
driver is restarting it after a failure, and from the kernel and applications, which continue to receive
service from the shadow.
Once the driver has restarted, the active-mode shadow returns the driver to its pre-failure state.
For example, the shadow re-establishes any configuration state and then replays pending requests.
59
A shadow driver is a “class driver,” aware of the interface to the drivers it shadows but not
of their implementations. The class orientation has two key implications. First, a single shadow
driver implementation can recover from a failure of any driver in its class, meaning that a handful
of different shadow drivers can serve a large number of device drivers. As previously mentioned,
Linux, for example, has only 20 driver classes. Second, implementing a shadow driver does not
require a detailed understanding of the internals of the drivers it shadows. Rather, it requires only
an understanding of those drivers’ interactions with the kernel. Thus, they can be implemented by
kernel developers with no knowledge of device specifics and have no dependencies on individual
drivers. For example, if a new network interface card and driver are inserted into a PC, the existing
network shadow driver can shadow the new driver without change. Similarly, drivers can be patched
or updated without requiring changes to their shadows.
5.2.2 Taps
As previously described, a shadow driver monitors communication between a functioning driver
and the kernel and impersonates one to the other during failure and recovery. This is made possible
by a new mechanism, called a tap. Conceptually, a tap is a T-junction placed between the kernel and
its drivers. During a shadow’s passive-mode operation, the tap: (1) invokes the original driver, and
then (2) invokes the corresponding shadow with the parameters and results of the call, as shown in
Figure 5.1.
When a failure is detected in a driver, its taps switch to active mode, as shown in Figure 5.2.
Here, both the kernel and the recovering device driver interact only with the shadow driver. Com-
munication between the driver and kernel is suspended until recovery is complete and the shadow
returns to passive mode, at which point the tap returns to its passive-mode state.
Taps depend on the ability to dynamically dispatch all communication between the driver and
the kernel. Consequently, all communication into and out of a shadowed driver must be explicit,
such as a procedure call or a message. Although most drivers operate this way, some do not. For
example, kernel-mode video drivers often communicate with user-mode applications through shared
memory regions [119]. Without additional support from the driver or its applications, these drivers
cannot be recovered without impacting applications.
60
Sound-Card Device Driver
Shadow Sound Driver
Kern
el
Inte
rface
Sound D
river
Cla
ss Inte
rface
OS Kernel
Kernel Interface
Sound DriverClass Interface
Sound Card
Taps
Copies
Figure 5.1: A sample shadow driver operating in passive mode.
Sound-Card Device Driver
Shadow Sound Driver
Kern
el
Inte
rface
Sound D
river
Cla
ss Inte
rface
OS Kernel
Kernel Interface
Sound DriverClass Interface
Sound Card
Taps
Figure 5.2: A sample shadow driver operating in active mode.
5.2.3 The Shadow Recovery Manager
The shadow recovery manager is responsible for coordinating recovery with shadow drivers. The
Nooks controller notifies the shadow recovery manager that it has detected a failure in a driver. The
shadow recovery manager then transitions the shadow driver to active mode and closes the taps. In
this way, requests for the driver’s services are redirected to the corresponding shadow driver. The
shadow recovery manager then initiates the shadow driver’s recovery sequence to restore the driver.
61
When recovery completes, the shadow recovery manager returns the shadow driver to passive-mode
operation and re-opens its taps so that the driver can resume service.
5.2.4 Summary
The shadow driver design simplifies the development and integration of shadow drivers into existing
systems. Each shadow driver is a single module written with knowledge of the behavior (interface)
of a class of device drivers, allowing it to conceal a driver failure and restart the driver after a
failure. A shadow driver, when passive, monitors communication between the kernel and the driver.
It becomes active when a driver fails and then both proxies requests for the driver and restores the
driver’s state.
5.3 Implementation
This section describes the implementation of shadow drivers in the Linux operating system. I have
implemented shadow drivers for three classes of device drivers: sound-card drivers, network inter-
face drivers, and IDE storage drivers.
5.3.1 General Infrastructure
Shadow drivers rely on a reliability layer to provide three services within the kernel. An isolation
service prevents driver errors from corrupting the kernel by stopping a driver upon detection of a
failure. An interposition service provides the taps required for transparent shadowing and recovery.
Lastly, an object tracking service tracks kernel resources associated with the driver to facilitate
reclamation during the driver’s recovery.
The Nooks controller, described in the previous chapter, handles the initial installation of shadow
drivers. The controller creates a new shadow driver instance when a driver is loaded into the kernel.
Because a single shadow driver services an entire class of device drivers, there may be several
instances of a shadow driver executing if there is more than one driver of a class present.
Shadow drivers and Nooks together form a driver recovery subsystem, as shown in Figure 5.3.
The Nooks reliability layer contributes lightweight kernel protection domains, wrappers, and the
62
Linux KernelRecovery Subsystem
ApplicationApplication
Nooks Reliability Layer
Shadow DriverShadow DriverShadow Recovery
Manager
Protection Domains
Shadow Driver
Application
Device
Device
Device
Device
Driver
Driver
Driver
DriverObject
DatabaseWrappers
Taps
Figure 5.3: The Linux operating system with several device drivers and the shadow driver recoverysubsystem.
object tracker’s database. The shadow driver recovery mechanism extends Nooks by incorporating
taps into the wrapper code, and by adding the shadow recovery manager and a set of shadow drivers.
5.3.2 Passive-Mode Monitoring
In passive mode, a shadow driver monitors the current state of a device driver by observing its
communication with the kernel. In order to return the driver to its state at the time of failure, the
shadow records the inputs to the driver in a log. These inputs are then replayed during recovery.
With no knowledge of how drivers operate, the shadow would have to log all inputs to the driver.
However, because the shadow is implemented with knowledge of the driver’s interface, and hence its
abstract state machine, not all inputs must be logged. Instead, the shadow only records inputs needed
to return a driver to its state at the time of failure. The shadow drops requests that do not advance
the driver’s state or whose impact has been superseded by later inputs, for example transitions on a
loop in the abstract state machine.
Figure 5.4 shows an example of the state machine transitions for a sound-card driver. The
transitions, made when the kernel issues requests to the driver, are numbered. The final state of the
sequence is S4, but there is a loop through state S3. As a result, the shadow may drop requests 2
through 5 from its log, because they do not affect the final state of the driver.
To implement this state machine, the shadow driver maintains a log in which it records several
63
S2
S3
1: open
2: ioctl-1
5:ioctl-reset
S4
4: ioctl-2
6: ioctl-1
3: write
S1
Figure 5.4: The State machine transitions for a sound-card shadow driver. Those recorded forrecovery are shown in boldface.
types of information. First, it tracks I/O requests made to the driver, enabling pending requests to
be re-submitted after recovery. An entry remains in the log until the corresponding request has been
handled. In addition, for connection-oriented drivers, the shadow driver records the state of each
active connection, such as offset or positioning information.
The shadow driver also records configuration and driver parameters that the kernel passes into
the driver. The shadow relies on this information to reconfigure the driver to its pre-failure state
during recovery. For example, the shadow sound-card driver logs ioctl calls (command numbers
and arguments) that configure the driver.
For stateful devices, such as a hard disk, the shadow does not create a copy of the device state.
Instead, a shadow driver depends on the fail-stop assumption to preserve persistent state (e.g., on
disk) from corruption. In other cases, the shadow may be able to force the device’s clients to recreate
the state after a failure. For example, a windowing system can recreate the contents of a frame buffer
by redrawing the desktop.
In many cases, passive-mode calls do no work and the shadow returns immediately to the caller.
For example, the kernel maintains a queue of outstanding requests to a disk driver, and hence the
shadow driver for an IDE disk does little in passive mode. For the network shadow driver, too,
the Nooks object-tracking system performs much of the work to capture driver state by recording
outstanding packets.
64
Opaque parameters can pose problems for recovery as they did for isolation. However, the class-
based approach allows shadow drivers to interpret most opaque pointers. The standardized interface
to drivers ensures that a client of the interface that has no knowledge of a driver’s implementation
can still invoke it. Hence, clients must know the real type of opaque parameters. The shadow
implementer can use the same knowledge to interpret them. For example, the Open Sound System
interface [197] defines opaque pointer parameters to the ioctl call for sound-card drivers. The
shadow sound-card driver relies on this standard to interpret and log ioctl requests.
5.3.3 Active-Mode Recovery
The shadow enters active mode when a failure is detected in a driver. A driver typically fails by
generating an illegal memory reference or passing an invalid parameter across a kernel interface.
Nooks’ failure detectors notice the failure and notify the controller, which in turn invokes the shadow
recovery manager. This manager immediately locates the corresponding shadow driver and directs
it to recover the failed driver. The shadow driver’s task is to restore the driver to the state it was
in at the time of failure, so it can continue processing requests as if it had never failed. The three
steps of recovery are: (1) stopping the failed driver, (2) reinitializing the driver from a clean state,
and (3) transferring relevant shadow driver state into the new driver. Unlike Nooks’ restart recovery
manager, a shadow driver does not completely unload the failed driver.
Stopping the Failed Driver
The shadow recovery manager begins recovery by informing the responsible shadow driver that
a failure has occurred. It also closes the taps, isolating the kernel and driver from one another’s
subsequent activity during recovery. After this point, the tap redirects all kernel requests to the
shadow until recovery is complete.
Informed of the failure, the shadow driver first invokes the isolation service to preempt threads
executing in the failed driver. It also disables the hardware device to prevent it from interfering
with the OS while not under driver control. For example, the shadow disables the driver’s interrupt
request line. Otherwise, the device may continuously interrupt the kernel and prevent recovery. On
hardware platforms with I/O memory mapping, the shadow also removes the device’s I/O mappings
65
to prevent DMAs into kernel memory.
In preparation for recovery, the shadow garbage collects resources held by the driver. To en-
sure that the kernel does not see the driver “disappear” as it is restarted, the shadow retains ob-
jects that the kernel uses to request driver services. For example, the shadow does not release the
network device object for network device drivers. The remaining resources, not needed for
recovery, are released.
Reinitializing the Driver
The shadow driver next “boots” the driver from a clean state. Normally, booting a driver requires
loading the driver from disk. However, the disk driver may not be functional during recovery.
Hence, the driver code and data must already be in memory before a failure occurs. For this reason,
the shadow caches a copy of the device driver’s initial, clean data section when the driver is first
loaded. These data sections tend to be small. The driver’s code is already loaded read-only in the
kernel, so it can be reused from memory.
The shadow boots the driver by repeating the sequence of calls that the kernel makes to initialize
a driver. For some driver classes, such as sound-card drivers, this consists of a single call into the
driver’s initialization routine. Other drivers, such as network interface drivers, require additional
calls to connect the driver into the network stack.
As the driver restarts, the shadow reattaches the driver to the kernel resources it was using before
the failure. For example, when the driver calls the kernel to register itself as a driver, the taps redirect
these calls to the shadow driver, which reconnects the driver to existing kernel data structures. The
shadow reuses the existing driver registration, passing it back to the driver. For requests that generate
callbacks, such as a request to register the driver with the PCI subsystem, the shadow emulates the
kernel and calls the driver back in the kernel’s place. The shadow also provides the driver with
its hardware resources, such as interrupt request lines and memory-mapped I/O regions. If the
shadow had disabled these resources in the first step of recovery, the shadow re-enables them, e.g.,
enabling interrupt handling for the device’s interrupt line. In essence, the shadow driver initializes
the recovering driver by calling and responding as the kernel would when the driver starts normally.
Unfortunately, not all drivers implement the driver interface specification correctly. For example,
66
some drivers depend on the exact timing of the calls made during initialization. While my shadow
network driver was able to restart five of the six network drivers that I tested, the e100 network
driver does not fully initialize during the open call as the specification requires. Instead, the driver
sets a timer that fires between the open call and the first call to send a packet. When a user-mode
program loads the driver normally, there is ample delay between these two calls for the timer to
fire. However, when the shadow driver invokes the driver during recovery, these calls come in rapid
succession. As a result, the driver has not finished initializing when the first packet arrives. This
causes the recovery process to fail because the shadow cannot replay the log of pending requests.
For testing purposes, I implemented a work-around by inserting a delay to allow any pending
timers to fire after calling the open function. However, this example raises the larger issue of
compatibility. Shadow drivers rely on driver writers to implement to the published interface and
not to the specifics of the kernel implementation. The implementer of a shadow driver may choose
simplicity, by only supporting those drivers that correctly implement the interface, or coverage, by
including extra code to more closely mimic the kernel’s behavior.
Transferring State to the New Driver
The final recovery step restores the driver to the state it was in at the time of the failure, permitting
it to respond to requests as if it had never failed. Thus, any configuration that either the kernel or
an application had downloaded to the driver must be restored. The shadow driver walks its log and
issues requests to the driver that to restore its state.
The details of this final state transfer depend on the device driver class. Some drivers are con-
nection oriented. For these, the state consists of the state of the connections before the failure. The
shadow re-opens the connections and restores the state of each active connection with configuration
calls. Other drivers are request oriented. For these, the shadow restores the state of the driver by
replaying logged configuration operations and then resubmits to the driver any requests that were
outstanding when the driver crashed.
As an example, to restart a sound-card driver, the shadow driver resets the driver and all its open
connections back to their pre-failure state. Specifically, the shadow scans its list of open connections
and calls the open function in the driver to reopen each connection. The shadow then walks its log
67
of configuration commands for each connection and replays commands that set driver properties.
When replaying commands, the shadow must ensure that parameters that usually have user-
mode addresses are handled correctly. Drivers normally return an error if parameters from user-
mode have kernel-mode addresses. For these calls, shadow drivers reset the user-mode address limit
for the current task to include the kernel address space. This ensures that the copy from user
and copy to user accessor functions accept kernel addresses. Thus, the driver is unaware that
the parameter is in the kernel rather than user space.
For some driver classes, the shadow cannot completely transfer its state into the driver. However,
it may be possible to compensate in other, perhaps less elegant, ways. For example, a sound-card
driver that is recording sound stores the number of bytes it has recorded since the last reset. After
recovery, the sound-card driver initializes this counter to zero. Because the interface has no call that
sets the counter value, the shadow driver must insert its “true” value into the return argument list
whenever the application reads the counter to maintain the illusion that the driver has not crashed.
The shadow can do this because it receives control (on its replicated call) before the kernel returns
to user space.
After resetting driver and connection state, the shadow must handle requests that were either
outstanding when the driver crashed or arrived while the driver was recovering. If a driver crashes
after submitting a request to a device but before notifying the kernel that the request has completed,
the shadow cannot know whether the device completed the request. As a result, shadow drivers
cannot guarantee exactly once behavior and must rely on devices and higher levels of software to
absorb duplicate requests. So, the shadow driver has two choices during recovery: restart in-progress
requests and risk duplication, or cancel the request and risk lost data. For some device classes, such
as disks or networks, duplication is acceptable. However, other classes, such as printers, may not
tolerate duplicates. In these cases, the shadow driver cancels outstanding requests and returns an
error to the kernel or application in a manner consistent with the driver interface. As I discuss when
evaluating shadow drivers in Chapter 6, this is not ideal but may be better than other alternatives,
such as crashing the system or the application.
After this final step, the driver has been reinitialized, linked into the kernel, reloaded with its
pre-failure state, and is ready to process commands. At this point, the shadow driver notifies the
shadow recovery manager, which sets the taps to restore kernel-driver communication and reestab-
68
lish passive-mode monitoring.
5.3.4 Active-Mode Proxying of Kernel Requests
While a shadow driver is restoring a failed driver, it is also acting as a proxy for the driver to
conceal the failure and recovery from applications and the kernel. Thus, the shadow must respond
to any request for the driver’s service in a way that satisfies and does not corrupt the driver’s caller.
The shadow’s response depends on the driver’s interface and the request semantics. In general, the
shadow will take one of five actions:
1. Respond with information that it has recorded in its log.
2. Report that the driver is busy and that the kernel or application should try again later.
3. Suspend the request until the driver recovers.
4. Queue the request for processing after recovery.
5. Silently drop the request.
The choice of strategy depends on the caller’s expectations of the driver.
Writing the proxying code requires knowledge of the kernel-driver interface, its interactions,
and its requirements. For example, the kernel may require that some driver functions never block,
while others always block. Some kernel requests are idempotent (e.g., many ioctl commands),
permitting duplicate requests to be dropped, while others return different results on every call (e.g.,
many read requests). The writer of a shadow for a driver class uses these requirements to select
the response strategy.
Device drivers often support the concept of being “busy.” This concept allows a driver to man-
age the speed difference between software running on the computer and the device. For example,
network drivers in Linux may reject requests and turn themselves off if packets are arriving from
the kernel to quickly and their queues are full. The kernel then refrains from sending packets until
the driver turns itself back on. The notion of being “busy” in a driver interface simplifies active
proxying. By reporting that the device is currently busy, shadow drivers instruct the kernel or appli-
cation to block calls to a driver. The shadow network driver exploits this behavior during recovery
by returning a “busy” error on calls to send packets. IDE storage drivers support a similar notion
when request queues are full. Sound drivers can report that their buffers are temporarily full.
69
Table 5.1: The proxying actions of the shadow sound-card driver.
Request Actionread / write suspend callerinterrupt drop requestquery capability ioctl answer from logquery buffer ioctl act busyreset ioctl queue for later / drop duplicate
The shadow sound-card driver uses a mix of all five strategies for proxying functions in its
service interface. Table 5.1 shows the shadow’s actions for common requests. The shadow sus-
pends kernel read and write requests, which play and record sound samples, until the failed
driver recovers. It processes ioctl calls itself, either by responding with information it captured
or by logging the request to be processed later. For ioctl commands that are idempotent, the
shadow driver silently drops duplicate requests. Finally, when applications query for buffer space,
the shadow responds that buffers are full. As a result, many applications block themselves rather
than blocking in the shadow driver.
5.3.5 Code Size
The preceding sections described the implementation of shadow drivers. Their class-based approach
allows a single shadow driver to recover from the failure of any driver in its class. Thus, the shadow
driver architecture leverages a small amount of code, in a shadow driver, to recover from the failure
of a much larger body of code, the drivers in the shadow’s class. However, if the shadow itself
introduces significant complexity, this advantage is lost. This section examines the complexity of
shadow drivers in terms of code size, which can serve as a proxy for complexity.
As I previously mentioned, I implemented shadow drivers for three classes of device drivers:
sound-card drivers, network interface drivers, and IDE storage drivers. Table 5.2 shows, for each
class, the size in lines of code of the shadow driver for the class. For comparison, I also show
the size of the drivers that I tested for Chapter 6, and the total number and cumulative size of all
Linux device drivers in the 2.4.18 kernel from each class. The total code size is an indication of the
leverage gained through the shadow’s class-driver structure. The table demonstrates that a shadow
70
Table 5.2: The size and quantity of shadows and the drivers they shadow.
Shadow Driver Device Driver Shadowed Class Size Class SizeDriver Class Lines of Code Lines of Code # of Drivers Lines of CodeSound 666 7,381 (audigy) 48 118,981Network 198 13,577 (e1000) 190 264,500Storage 321 5,358 (ide-disk) 8 29,000
driver is significantly smaller than the device driver it shadows. For example, the sound-card shadow
driver is only 9% of the size of the audigy device driver it shadows. The IDE storage shadow is only
6% percent of the size of the Linux ide-disk device driver.
On top of the 26,000 lines of code in Nooks, I added about 3300 lines of new code to implement
shadow drivers. I made no additional changes to the Linux kernel. Beyond the 1200 lines of code
in the shadow drivers themselves, additional code consisted of approximately 600 lines of code for
the shadow recovery manager, 800 lines of common code shared by all shadow drivers, and another
750 lines of code for general utilities, such as tools to automatically generate the tap code. Of the
177 taps I inserted, only 31 required actual code in a shadow; the remainder were no-ops because
the calls did not significantly impact kernel or driver state.
5.3.6 Limitations
As previously described, shadow drivers have limitations. Most significant, shadow drivers rely on
a standard interface to all drivers in a class. If a driver provides “enhanced” functionality with new
methods, the shadow for its class will not understand the semantics of these methods. As a result, it
cannot replicate the driver’s state machine and recover from its failure.
Shadow drivers rely on explicit communication between the device driver and kernel. If kernel-
driver communication takes place through an ad-hoc interface, such as shared memory, the shadow
driver cannot monitor it. Furthermore, shadow drivers require that all “important” driver state is
visible across the kernel-driver interface. The shadow driver cannot restore internal driver state
derived from device inputs. If this state impacts request processing, the shadow may be unable to
restore the driver to its pre-failure state. For example, a TCP-offload engine that stores connection
71
state cannot be recovered, because the connection state is not visible on the kernel-driver interface.
While shadow drivers are built on top of Nooks, they can be used in other contexts where Nooks’
services are available. Microkernels provide much of this infrastructure and could be anther platform
for shadow drivers [97]. However, the capabilities of the underlying isolation system limit the
effectiveness of shadow drivers. If the isolation system cannot prevent kernel corruption, shadow
drivers cannot facilitate system recovery. Shadow drivers also assume that driver failures do not
cause irreversible side effects. If a corrupted driver stores persistent state (e.g., printing a bad check
or writing bad data on a disk), the shadow driver will not be able to correct that action.
In addition, shadow drivers rely on the underlying isolation system to detect failures. If it does
not, then shadow drivers will not recover and applications may fail. Detecting failures is difficult
because drivers are complex and may respond to application requests in many ways. It may be
impossible to detect a valid but incorrect return value; for example, a sound-card driver may return
incorrect sound data when recording. As a result, no failure detector can detect every device driver
failure. However, shadows could provide class-based failure detectors that detect violations of a
driver’s programming interface and reduce the number of undetected failures.
Shadow drivers can only recover from transient failures because deterministic failures may recur
during recovery. While unable to recover, shadow drivers are still useful for these failures. The
sequence of shadow driver recovery events creates a detailed reproduction scenario of the failure
that aids diagnosis. This record contains the driver’s calls into the kernel, requests to configure the
driver, and I/O requests that were pending at the time of failure. This information enables a software
engineer to find and fix the offending bug more efficiently.
Finally, shadow drivers may not be suitable for applications with real-time demands. During
recovery, a device may be unavailable for several seconds without notifying the application of a
failure. These applications, which should be written to tolerate failures, would be better served by a
solution that restores the driver state but does not perform active proxying.
5.4 Summary
Improving the reliability of commodity systems demands that device driver failures be tolerated.
While Nooks was designed to keep the OS running in the presence of driver failure, it did little for
72
applications that depend on a driver. To improve application reliability, I designed and implemented
shadow drivers. Shadow drivers mask device driver failures from both the OS and applications.
Shadow drivers provide an elegant mechanism that leverages the properties of device drivers for
recovery. Based on an abstract state machine modeling an entire class of drivers, shadow drivers
monitor the communication between the kernel and driver to obtain the driver state during normal
operation. When a driver fails, the shadow relies on this state to conceal the failure by proxying
for the driver. At the same time, the shadow recovers by restarting the driver, and then replaying
requests to bring the driver back to its pre-failure state.
My experience shows that shadow drivers, while simple to implement and integrate, are highly
leveraged: a single shadow driver can enable recovery for an entire class of device drivers. Thus,
with the addition of a small amount of shadow code to an operating system, the failure of a much
larger body of driver code can be tolerated. In Chapter 6 I show that shadow drivers can keep
applications running in the presence of failures, while Nooks’ recovery managers cannot.
73
Chapter 6
EVALUATION OF NOOKS AND SHADOW DRIVERS
6.1 Introduction
The thesis of my work is that Nooks and shadow drivers can significantly improve system reliability
by protecting the kernel and applications from driver failures. This chapter addresses four key
questions:
1. Isolation. Can the operating system survive driver failures and crashes?
2. Recovery. Can applications that use a device driver continue to run even after the driver fails?
3. Assumptions. How reasonable is the assumption that driver failures are fail-stop when isolated
by Nooks?
4. Performance. What is the performance overhead of Nooks in the absence of failures, i.e., the
steady-state cost?
Using synthetic fault injection experiments, I show that Nooks can isolate and automatically recover
from faults in drivers. In these tests, the operating system survived 99% of driver faults that would
otherwise crash Linux. Furthermore, shadow drivers were able to restart the driver while keeping
applications running in 98% of failures. I also show that the majority of driver failures are fail-stop
under Nooks, and that that Nooks is limited in this regard by its ability to detect failures and not by
its ability to isolate drivers. Finally, I show that when running common applications, Nooks imposes
a low to moderate performance overhead.
6.2 Test Methodology
6.2.1 Types of Extensions Isolated
In the experiments reported below, I used Nooks to isolate three types of extensions: device drivers,
a kernel subsystem (VFAT), and an application-specific kernel extension (kHTTPd). The device
74
drivers I chose were common network, sound-card, and IDE storage drivers, representative of three
major driver interfaces (character, network, and block [52]).
A device driver’s interaction with the kernel is well matched to the Nooks isolation model for
many reasons. First, the kernel invokes drivers through narrow, well-defined interfaces; therefore,
it is straightforward to design and implement their wrappers. Second, drivers frequently deal with
blocks of opaque data, such as network packets or disk blocks, that do not require validation. Third,
drivers often batch their processing to amortize interrupt overheads. When run with Nooks, batching
also reduces isolation overhead.
While Nooks is designed primarily for device drivers, it can also be used to tolerate the failure of
other kernel extensions. As a demonstration, I tested Nooks with two non-driver kernel extensions.
First, I isolated a loadable kernel subsystem. The subsystem I chose was the optional VFAT file
system, which is compatible with the Windows 95 FAT32 file system [144]. While drivers tend to
have a small number of interfaces with relatively few functions, the file system interface is larger and
more complex. VFAT has six distinct interfaces that together export over 35 calls; by comparison, the
device drivers each have one or two interfaces with between six and thirteen functions. In addition,
driver interfaces tend to pass relatively simple data structures, such as network packets and device
objects, while the file system interfaces pass complex, heavily linked data structures such as inodes.
Second, I isolated an application-specific kernel extension – the kHTTPd Web server [200].
kHTTPd resides in the kernel so that it can access kernel network and file system data structures
directly, avoiding otherwise expensive system calls. My experience with kHTTPd demonstrates that
Nooks can isolate even ad-hoc and unanticipated kernel extensions.
Overall, I have isolated fourteen drivers and two other kernel extensions under Nooks, as shown
in Table 6.1. I present reliability and performance results for seven of the extensions representing
the three driver classes and the other extension types: sb, audigy, pcnet32, e1000, ide-disk, VFAT
and kHTTPd. Results for the remaining drivers are consistent with those presented.
6.2.2 Test Platforms
To evaluate Nooks isolation and shadow driver recovery, I produced three OS configurations based
on the Linux 2.4.18 kernel:
75
Table 6.1: The Linux extensions tested. Results are reported for those in boldface.
Class Driver DeviceNetwork e1000 Intel Pro/1000 Gigabit Ethernet
pcnet32 AMD PCnet32 10/100 Ethernet3c59x 3COM 3c509b 10/100 Ethernet3c90x 3COM 3c90x series 10/100 Ethernet drivere100 Intel Pro/100 Ethernetepic100 SMC EtherPower 10/100 Ethernet
Sound audigy SoundBlaster Audigy sound cardsb SoundBlaster 16 sound cardemu10k1 SoundBlaster Live! sound cardes1371 Ensoniq sound cardcs4232 Crystal sound cardi810 audio Intel 810 sound card
Storage ide-disk IDE diskide-cd IDE CD-ROM
None VFAT Win95 compatible file systemkHTTPd In-kernel Web server
1. Linux-Native is the unmodified Linux kernel.
2. Linux-Nooks is a version of Linux-Native that includes the Nooks fault isolation subsystem
and the restart recovery manager. When a driver fails, this system restarts the driver but does
not attempt to conceal its failure.
3. Linux-SD includes Nooks, the shadow driver recovery manager, and the three shadow drivers.
For isolation tests, I use Linux-Nooks, as it can restart both device drivers and other kernel
extensions. For the recovery tests, I use Linux-SD.
6.2.3 Fault Injection
Ideally, I would test Nooks with a variety of real drivers long enough for many failures to occur.
This is not practical, however, because of the time needed for testing. For example, with a mean
time-to-failure of 1 week, it would take almost 10 machine years to observe 500 failures. Instead, I
use synthetic fault injection to insert artificial faults into drivers. I adapted a fault injector developed
76
Table 6.2: The types of faults injected into extensions.
Fault Type Code TransformationSource fault Change the source registerDestination fault Change the destination registerPointer fault Change the address calculation for a memory instructionInterface fault Use existing value in register instead of passed parameterBranch fault Delete a branch instructionLoop fault Invert the termination condition of a loop instructionText fault Flip a bit in an instructionNOP fault Elide an instruction
for the Rio File Cache [157] and ported it to Linux. The injector automatically changes single
instructions in driver code to emulate a variety of common programming errors, such as uninitialized
local variables, bad parameters, and inverted test conditions.
I inject two different classes of faults into drivers. First, I inject faults that emulate specific pro-
gramming errors that are common in kernel code [191, 46]. Source and destination faults emulate
assignment errors by changing the operand or destination of an instruction. Pointer faults emulate
incorrect pointer calculations and cause memory corruption. Interface faults emulate bad parame-
ters. I emulate bugs in control flow through branch faults, which remove a branch instruction, and
by loop faults, which change the termination condition for a loop.
Second, I expand the range of testing by injecting random changes that do not model specific
programming errors. In this category are text faults, in which I flip a random bit in a random
instruction, and NOP faults, in which I delete a random instruction. Table 6.2 shows the types of
faults I inject, and how the injector simulates programming errors (see [157] for a more complete
description of the fault injector).
Unfortunately, there is no available information on the distribution of programming errors for
real-world driver failures. Therefore, I inject the same number of faults of each type. The output
of the fault injection tests is therefore a metric of coverage, not reliability. The tests measure the
fraction of faults (failure causes) that can be detected and isolated, not the fraction of existing failures
that can be tolerated.
77
6.3 System Survival
In this section, I evaluate Nooks’ ability to isolate the kernel from extension failure. The goal of
these tests is to measure the survival rate of the operating system. I consider application failures in
Section 6.4.
6.3.1 Test Environment
The application-level workload consists of four programs that stress the sound-card driver, the net-
work driver, the storage driver and VFAT, and kHTTPd. The first program plays a short MP3 file.
The second performs a series of ICMP-ping and TCP streaming tests, while the third untars and
compiles a number of files. The fourth program runs a Web load generator against the kernel-mode
Web server.
To ensure a common platform for testing with both the device drivers and the other extensions,
I use the restart recovery manager to restore extension operation after a failure. There is one excep-
tion: the tests for ide-disk use shadow drivers for recovery. The restart recovery manager depends
on accessing the disk to reload a failed driver. Hence, when the disk driver fails, the restart manager
is unable to reload it. In contrast, the shadow driver recovery manager keeps the driver image in
memory and does not need to reload it off disk.
I ran most isolation experiments in the context of the VMware Virtual Machine [190]. The
virtual machine allows me to perform thousands of tests remotely while quickly and easily returning
the system to a clean state following each one. I spot-checked a number of the VMware trials against
a base hardware configuration (i.e., no virtual machine) and discovered no anomalies. The e1000
tests, however, were run directly on raw hardware, because VMware does not emulate the Intel
Pro/1000 Gigabit Ethernet card.
To measure isolation, I conducted a series of trials in which I injected faults into extensions
running under two different Linux configurations, both running the Linux-Nooks kernel. In the
first, called “native,” the Nooks isolation services were present but unused.1 In the second, called
“Nooks,” the isolation services were enabled for the extension under test. For each extension, I ran
1To ensure that I injected exactly the same fault in both systems requires using the same binary module. Linux doesnot support binary compatibility for loadable modules across different kernel versions. Hence, I used the same kernelfor both tests.
78
400 trials (50 of each fault type) on the native configuration. In each trial, I injected five random
faults into the extension and exercised the system, observing the results. I then ran those same
400 trials, each with the same five faults, against Nooks. It is important to note that the native
and Nooks configurations are identical binaries, allowing the automatic fault injector to introduce
identical faults.
6.3.2 System Survival Test Results
As described above, I ran 400 fault-injection trials for each of the six measured extensions on the
native and Nooks configurations. Not all fault-injection trials cause faulty behavior, e.g., bugs in-
serted on a rarely (or never) executed path will rarely (or never) produce an error. However, many
trials do cause failures. I now examine different types of failures that occurred.
System Crashes
A system crash is the most extreme and easiest problem to detect, as the operating system panics,
becomes unresponsive, or simply reboots. In an ideal world, every system crash caused by a fault-
injection trial under native Linux would result in a recovery under Nooks. As previously discussed,
in practice Nooks may not detect or recover from certain failures caused by very bad programmers
or very bad luck.
Figure 6.1 shows the number of system crashes caused by the fault-injection experiments for
each of the extensions running on native Linux and Nooks. Of the 517 crashes observed with native
Linux, Nooks eliminated 512, or 99%. In the remaining five crashes, the system deadlocked when
the driver went into a tight loop with interrupts disabled. Nooks does not detect this type of failure.
Figure 6.1 also illustrates a substantial difference in the number of system crashes that occur
for VFAT and sb extensions under Linux, compared to e1000, pcnet32, ide-disk and kHTTPd. This
difference reflects the way in which Linux responds to kernel failures. The e1000, pcnet32 and
ide-disk extensions are interrupt oriented, i.e., kernel-mode extension code is run as the result of an
interrupt. VFAT and sb extensions are process oriented, i.e., kernel-mode extension code is run as
the result of a system call from a user process. kHTTPd is process oriented but manipulates (and
therefore can corrupt) interrupt-level data structures. Linux treats exceptions in interrupt-oriented
79
System Crashes
10
52
119
152
9
175
1 0 0 0 2 2
0
40
80
120
160
200
sb e1000 pcnet32 ide-disk VFAT kHTTPd
Extension under test
Nu
mb
er
of
cra
sh
es
Native Nooks
Figure 6.1: The reduction in system crashes observed using Nooks.
code as fatal and crashes the system, hence the large number of crashes in e1000, pcnet32, ide-
disk and kHTTPd. Linux treats exceptions in process-oriented code as non-fatal, continuing to run
the kernel but terminating the offending process even though the exception occurred in the kernel.
This behavior is unique to Linux. Other operating systems, such as Microsoft Windows XP, deal
with kernel processor exceptions more aggressively by always halting the operating system. In such
systems, exceptions in VFAT and sb would cause system crashes.
Non-Fatal Extension Failures
While Nooks is designed to protect the OS from misbehaving extensions, it is not designed to detect
all erroneous extension behavior. For example, the network could disappear because the device
driver corrupts the device registers, or a mounted file system might simply become non-responsive
due to a bug. Neither of these failures is fatal to the system in its own right, and Nooks generally
does not detect such problems (nor is it intended to). However, when Nooks’ simple failure detectors
do detect such problems, its recovery services can safely restart the faulty extensions.
80
Non-fatal Extension Failures
156
17
33
0
232
33
56
16
31
80
27
0
50
100
150
200
250
sb e1000 pcnet32 ide-disk VFAT kHTTPd
Extension under test
Nu
mb
er o
f fa
ilu
res
Native Nooks
Figure 6.2: The reduction in non-fatal extension failures observed using Nooks.
The fault-injection trials cause a number of non-fatal extension failures, allowing me to examine
Nooks’ effectiveness in dealing with these cases, as well. Figure 6.2 shows the extent to which
Nooks reduces non-fatal extension failures that occurred in native Linux. In reality, these results
are simply a reflection of the Linux handling of process- and interrupt-oriented extension code, as
described earlier. That is, Nooks can trap exceptions in process-oriented extensions and can recover
the extensions to bring them to a clean state in many cases.
For the two interrupt-oriented Ethernet drivers (e1000 and pcnet32), Nooks already eliminated
all system crashes resulting from extension exceptions. The remaining non-crash failures are those
that leave the device in a non-functional state, e.g., unable to send or receive packets. Nooks cannot
remove these failures for e1000 and pcnet32, since it cannot detect them. The few extension failures
it eliminated occurred when process-oriented code was manipulating the driver.
For the ide-disk storage driver, there were no non-fatal extension failures. The operating system
depends on the ide-disk to perform paging. The operating system crashes when it malfunctions,
because the driver can no longer swap pages to disk. As a result, all failures of ide-disk are system
81
crashes.
For VFAT and the sb sound card driver, Nooks reduced the number of non-fatal extension fail-
ures. These failures were caused by kernel exceptions in process-oriented code, which caused Linux
to terminate the calling process and leave the extension in an ill-defined state. Nooks detected
the processor exceptions and performed an extension recovery, thereby allowing the application to
continue, but without access to the failed extension. The remaining non-fatal extension failures,
which occurred under native Linux and Nooks, were serious enough to leave the extension in a non-
functioning state but not serious enough to generate a processor exception that could be trapped by
Nooks.
The kHTTPd extension is similar to the interrupt-oriented drivers because it causes corruption
that leads to interrupt-level faults. However, a small number of injected faults caused exceptions
within the kHTTPd process-oriented code. Nooks caught these exceptions and avoided an extension
failure.
In general, the remaining non-fatal extension failures under Nooks were the result of deadlock
or data structure corruption within the extension itself. Fortunately, such failures were localized to
the extension and could usually be recovered from once discovered. In Section 6.5 I further analyze
the need for better error detection that is attuned to a specific class of extensions.
Manually Injected Faults
In addition to the automatic fault-injection experiments, I manually inserted 10 bugs across the
extensions. To generate realistic synthetic driver bugs, I analyzed patches posted to the Linux Kernel
Mailing List [133]. I found 31 patches that contained the strings “patch,” “driver,” and “crash” or
“oops” (the Linux term for a kernel fault) in their subject lines. Of the 31 patches, I identified 11 that
fix transient bugs (i.e., bugs that occur occasionally or only after a long delay from the triggering
test). The most common cause of failure (three instances) was a missing check for a null pointer,
often with a secondary cause of missing or broken synchronization. I also found missing pointer
initialization code (two instances) and bad calculations (two instances) that led to endless loops and
buffer overruns.
I “broke” extensions by removing checks for NULL pointers, failing to initialize stack and heap
82
variables properly, dereferencing a user-level pointer, and freeing a resource multiple times. Nooks
automatically detected and recovered from all such failures in all extensions.
Latent Bugs
Nooks revealed several latent bugs in existing kernel extensions. For example, it uncovered a bug
in the 3COM 3c90x Ethernet driver that occurs during its initialization.2 Nooks also revealed a bug
in another extension, kHTTPd [200], where an already freed object was referenced. In general, I
found that Nooks was a useful kernel development tool that provides a “fast restart” whenever an
extension under development fails.
6.3.3 Summary of System Survival Experiments
In testing its impact on operating system reliability, Nooks eliminated 99% of the system crashes
that occurred with native Linux. The remaining failures directly reflect the best-efforts principle
and are the cost, in terms of reliability, of an approach that imposes reliability on legacy extension
and operating systems code. In addition to crashes, Nooks can isolate and recover from many non-
fatal extension failures. While Nooks cannot detect many kinds of erroneous behavior, it can often
trap extension exceptions and initiate recovery. Overall, Nooks eliminated nearly 60% of non-fatal
extension failures in the fault-injection trials. Finally, Nooks detected and recovered from all of the
commonly occurring faults that I injected manually.
6.4 Application Survival
The previous section evaluated the ability of the operating system to survive extension failures. In
this section, I focus on applications. I answer the question of whether applications that use a device
driver continue to run even after the driver fails and recovers. I evaluate shadow driver recovery in
the presence of simple failures to show the benefits of shadow drivers compared to a system that
reveals failures to applications and the OS.
2If the driver fails to detect the card in the system, it immediately frees a large buffer. Later, when the driver isunloaded, it zeroes this buffer. Nooks caught this bug because it write-protected the memory when it was freed.
83
Table 6.3: The applications used for evaluating shadow drivers.
Device Driver Application ActivitySound • mp3 player (zinf) playing 128kb/s audio
(audigy driver) • audio recorder (audacity) recording from microphone• speech synthesizer (festival) reading a text file• strategy game (Battle of Wesnoth)
Network • network file transfer (scp) of a 1GB file• remote window manager (vnc)• network analyzer (ethereal) sniffing packets
Storage • compiler (make/gcc) compiling 788 C files(ide-disk driver) • encoder (LAME) converting 90 MB file .wav to .mp3
• database (mySQL) processing the Wisconsin Benchmark
6.4.1 Application Test Methodology
Unlike the system survival experiments, the application survival experiments use a wider variety
of drivers than VMware supports and were therefore run on a bare machine, a 3 GHz Pentium 4
PC with 1 GB of RAM and an 80 GB, 7200 RPM IDE disk drive. I built and tested three Linux
shadow drivers for three device-driver classes: network interface controller, sound card, and IDE
storage device. To ensure that my generic shadow drivers worked consistently across device driver
implementations, I tested them on fourteen different Linux drivers, shown in Table 6.1.3 Although
I present detailed results for only one driver in each class (e1000, audigy, and ide-disk), behavior
across all drivers was similar.
The crucial question for shadow drivers is whether an application can continue functioning fol-
lowing the failure of a device driver on which it relies. To answer this question, I tested the 10
applications shown in Table 6.3 on each of the three configurations, Linux-Native, Linux-Nooks, and
Linux-SD.
I simulated common bugs by injecting a software fault into a device driver while an application
using that driver was running. Because both Linux-Nooks and Linux-SD depend on the same isola-
tion and failure-detection services, I differentiate their recovery capabilities by simulating failures
that are easily isolated and detected. Based on the analysis of common drivers bugs in Section 6.3.2,
3Shadow drivers target device drivers, so I did not test the VFAT and kHTTPd kernel extensions.
84
Table 6.4: The observed behavior of several applications following the failure of the device driverson which they depend.
Application BehaviorDevice Driver Application Activity Linux-Native Linux-Nooks Linux-SD
Sound mp3 player CRASH MALFUNCTION√
(audigy driver) audio recorder CRASH MALFUNCTION√
speech synthesizer CRASH√ √
strategy game CRASH MALFUNCTION√
Network network file transfer CRASH√ √
(e1000 driver) remote window manager CRASH√ √
network analyzer CRASH MALFUNCTION√
IDE compiler CRASH CRASH√
(ide-disk driver) encoder CRASH CRASH√
database CRASH CRASH√
I injected a null-pointer dereference bug into my three drivers.
6.4.2 Application Test Results
Table 6.4 shows the three application behaviors I observed. When a driver failed, each application
continued to run normally (√
), failed completely (“CRASH”), or continued to run but behaved
abnormally (“MALFUNCTION”). In the latter case, manual intervention was typically required to
reset or terminate the program.
This table demonstrates that shadow drivers (Linux-SD) enable applications to continue running
normally even when device drivers fail. In contrast, all applications on Linux-Native failed when
drivers failed. Most programs running on Linux-Nooks failed or behaved abnormally, illustrating
that restart recovery protects the kernel, which is constructed to tolerate driver failures, but does not
protect applications. The restart recovery manager lacks two key features of shadow drivers: (1) it
does not advance the driver to its pre-fail state, and (2) it has no component to “pinch hit” for the
failed driver during recovery. As a result, Linux-Nooks handles driver failures by returning an error
to the application, leaving it to recover by itself. Unfortunately, few applications can do this.
Some applications on Linux-Nooks survived the driver failure but in a degraded form. For ex-
ample, mp3 player, audio recorder and strategy game continued running, but they lost their ability
85
to input or output sound until the user intervened. Similarly, network analyzer, which interfaces di-
rectly with the network device driver, lost its ability to receive packets once the driver was reloaded.
A few applications continued to function properly after driver failure on Linux-Nooks. One
application, speech synthesizer, includes the code to reestablish its context within an unreliable
sound-card driver. Two of the network applications survived on Linux-Nooks because they access
the network device driver through kernel services (TCP/IP and sockets) that are themselves resilient
to driver failures.
Unlike Linux-Nooks, Linux-SD can recover from disk driver failures. 4 Recovery is possible
because the IDE storage shadow driver instance maintains the failing driver’s initial state. During
recovery the shadow copies back the driver’s initial data and reuses the driver code, which is already
stored read-only in the kernel. In contrast, Linux-Nooks illustrates the risk of circular dependencies
from rebooting drivers. Following these failures, the restart recovery manager, which had unloaded
the ide-disk driver, was then required to reload the driver off the IDE disk. The circularity could
only be resolved by a system reboot. While a second (non-IDE) disk would mitigate this problem,
few machines are configured this way.
In general, programs that directly depend on driver state but are unprepared to deal with its
loss benefit the most from shadow drivers. In contrast, those that do not directly depend on driver
state or are able to reconstruct it when necessary benefit the least. My experience suggests that few
applications are as fault-tolerant as speech synthesizer. Were future applications to be pushed in
this direction, software manufacturers would either need to develop custom recovery solutions on a
per-application basis or find a general solution that could protect any application from the failure of
a device driver. Cost is a barrier to the first approach. Shadow drivers are a path to the second.
6.4.3 Application Behavior During Driver Recovery
Although shadow drivers can prevent application failure, they are not “real” device drivers and do
not provide complete device services. As a result, I often observed a slight timing disruption while
the driver recovered. At best, output was queued in the shadow driver or the kernel. At worst, input
was lost by the device. The length of the delay depends on the recovering device driver itself, which,
4For this reason, the ide-disk experiments in Section 6.3.2 use the shadow driver recovery manager.
86
on initialization, must first discover and then configure the hardware.
Few device drivers implement fast reconfiguration, which can lead to brief recovery delays.
For example, the temporary loss of the e1000 network device driver prevented applications from
receiving packets for about five seconds.5 Programs using files stored on the disk managed by the
ide-disk driver stalled for about four seconds during recovery. In contrast, the normally smooth
sounds produced by the audigy sound-card driver were interrupted by a pause of about one-tenth of
one second, which sounded like a slight click in the audio stream.
Of course, the significance of these delays depends on the application. Streaming applications
may become unacceptably “jittery” during recovery. Those processing input data in real-time might
become lossy. Others may simply run a few seconds longer in response to a disk that appears to be
operating more sluggishly than usual. In any event, a short delay during recovery is best considered
in light of the alternative: application and system failure.
6.4.4 Summary of Application Survival Experiments
This section examined the ability of shadow drivers to conceal failures from applications. The
results demonstrate that applications that failed in any form on Linux-Native or Linux-Nooks ran
normally with shadow drivers. Applications that depend on state in a driver benefited the most. In
all cases, recovery proceeded quickly, compared to a full system reboot.
6.5 Testing Assumptions
The preceding section assumed that failures were fail-stop. However, driver failures experienced in
deployed systems may exhibit a wider variety of behaviors. For example, a driver may corrupt the
application, kernel, or device without the failure being detected. In this situation, shadow drivers
may not be able to recover or mask failures from applications. This section uses fault injection
experiments in an attempt to generate faults that may not be fail-stop.
5This driver is particularly slow at recovery. The other network drivers I tested recovered in less than a second.
87
6.5.1 Testing the Fail-stop Assumption
If Nooks’ isolation services cannot ensure that driver failures are fail stop, then shadow drivers may
not be useful. To evaluate whether device driver failures can be made fail-stop, I performed large-
scale fault-injection tests of drivers and applications running on Linux-SD. For each driver and appli-
cation combination, I ran 400 fault-injection trials using the methodology from Section 6.2.3. How-
ever, because VMware does not support the drivers in these tests, they were run on real hardware
instead of a virtual machine. In total, I ran 2400 trials across the three drivers and six applications.
I ensured the errors were transient by removing them during recovery. After injection, I visually
observed the effect on the application and the system to determine whether a failure or recovery had
occurred. For each driver, I tested two applications with significantly different usage scenarios. For
example, I chose one sound-playing application (mp3 player) and one sound-recording application
(audio recorder).
If I observed a failure, I then assessed the trial on two criteria: whether Nooks detected the
failure, and whether the shadow driver could mask the failure and subsequent recovery from the
application. For undetected failures, I triggered recovery manually. Note that a user may observe a
failure that an application does not, for example, by testing the application’s responsiveness.
Figure 6.3 shows the results of my experiments. For each application, I show the percentage
of failures that the system detected and recovered automatically, the percentage of failures that the
system recovered successfully after manual failure detection, and the percentage of failures in which
recovery was not successful. The total number of failures that each application experienced is shown
at the bottom of the chart. Only 16% of the injected faults caused a failure.
In my tests, 390 failures occurred across all applications. The system automatically detected
65% of the failures. In every one of these cases, shadow drivers were able to mask the failure
and recover. Similar to the non-fatal extension failures when testing isolation, the system failed
to detect 35% of the failures. However, in these cases, I manually triggered recovery. Shadow
drivers recovered from nearly all of these failures (127 out of 135). Recovery was unsuccessful in
the remaining 8 cases because either the system had crashed (5 cases) or the driver had corrupted
the application beyond the possibility of recovery (3 cases). It is possible that recovery would have
succeeded had a better failure detector discovered these failures earlier.
88
Fault Injection Outcomes
0%
20%
40%
60%
80%
100%
mp3 player audio
recorder
network
file
transfer
network
analyzer
compiler database
Percen
t o
f O
utc
om
es
Automatic Recovery Manual Recovery Failed Recovery
Sound Network IDE
38 58974479 76
Figure 6.3: Results of fault-injection experiments on Linux-SD.
Across all applications and drivers, I found three major causes of undetected failure. First, the
system did not detect application hangs caused by I/O requests that never completed (omission fail-
ures). Second, the system did not detect errors in the interactions between the device and the driver,
e.g., incorrectly copying sound data to a sound card (output failures). Third, the system did not de-
tect certain bad parameters, such as incorrect result codes or data values (again, output failures). For
example, the ide-disk driver corrupted the kernel’s disk request queue with bad entries. Detecting
these three error conditions would require that the system better understand the semantics of each
driver class. For example, 68% of the sound driver failures with audio recorder went undetected.
This application receives data from the driver in real time and is highly sensitive to driver output. A
small error or delay in the results of a driver request may cause the application to stop recording or
record the same sample repeatedly.
My results demonstrate a need for better failure detectors to achieve high levels of reliability.
Currently, Nooks relies on generic failure detectors that are not specialized to the behavior of a
class of drivers. Shadow drivers provide an opportunity to add specialized detectors as part of the
passive-mode monitoring code.
89
6.5.2 Recovery Errors
The fault injection tests in Section 6.3.2 provide further data on whether Nooks can ensure that
extension failures are fail-stop. The VFAT file system manages persistent state stored on disk, and
hence there is some chance that the extension damaged critical on-disk structures before Nooks
detected an error condition.
In practice, I found that in 90% of the cases, VFAT recovery with the restart recovery manager
resulted in on-disk corruption (i.e., lost or corrupt files or directories). Since fault injection occurs
after many files and directories have been created, the abrupt shutdown and restart of the file system
leaves them in a corrupted state. As an experiment, I caused Nooks to synchronize the disks with
the in-memory disk cache before releasing resources on a VFAT recovery. This reduced the number
of corruption cases from 90% to 10%. These statistics are similar to what occurs when the system
crashes, and suggest that a recovery manager for file systems should perform this check during
recovery.
6.5.3 Summary
This section examined the assumption that Nooks can make driver failures fail-stop. My test results
show that most failures are fail-stop, but a substantial fraction are not. The problem was not one
of isolation, as the kernel and applications continued to execute after manually triggered recovery,
but of failure detection. Thus, further improving reliability demands driver class-specific failure
detectors to raise the detection rate.
6.6 Performance
This section presents benchmark results that evaluate the performance cost of the Nooks and shadow
drivers. The experiments use existing benchmarks and tools to compare the performance of a system
using Nooks to one that does not. I ran most tests on a Dell 3 GHz Pentium 4 PC running Linux
2.4.18. The machine includes 1 GB of RAM, a SoundBlaster Audigy sound card, an Intel Pro/1000
Gigabit Ethernet adapter, and a single 7200 RPM, 80 GB IDE hard disk drive. My network tests
90
Table 6.5: The relative performance of Nooks compared to native Linux for six benchmark tests.An ’*’ in the first column indicates a different hardware platform.
Benchmark Extension XPC Nooks Native NooksRate Relative CPU CPU
(per sec) Perf. (%) Util. (%) Util. (%)
mp3 player audigy 462 100 0.1 0.1audio recorder audigy 1059 100 0.3 0.6
network receive e1000 (receiver) 13,422 100 24.8 31.2network send e1000 (sender) 58,373 99 27.8 56.7DB queries ide-disk 294 97 21.4 26.9
Compile-local ide-disk 58 98 99.2 99.6Compile-local* VFAT 26,979 89 88.7 88.1
Serve-simple-web-page* kHTTPd (server) 61,183 44 96.6 96.8Serve-complex-web-page* e1000 (server) 1,960 97 90.5 92.6
used two identically equipped machines.6 A few tests, indicated by a ’*’ in Table 6.5, were run
on similar machines with a 1.7 GHz Pentium 4 with 890 MB of RAM. Unlike the reliability tests
described previously, I ran all performance tests on a bare machine, i.e., one without the VMware
virtualization system.
Table 6.5 summarizes the benchmarks used to evaluate system performance. For each bench-
mark, I used Nooks to isolate a single extension, indicated in the second column of the table. I
ran each benchmark on native Linux without Nooks (Linux-Native) and then again on a version of
Linux with Nooks and shadow drivers enabled (Linux-SD). I use the shadow driver recovery man-
ager for device drivers and the restart recovery manager for VFAT and kHTTPd. 7 The table shows
the relative change in performance for Nooks, either in wall clock time or throughput, depending
on the benchmark. I also show CPU utilization measured during benchmark execution (which is
only accurate to a few percent), as well as the rate of XPCs per second incurred during each test. I
compute relative performance either by comparing latency (Play-mp3, Compile-local, DB queries)
or throughput (network send, network receive, Serve-simple-web-page, Serve-complex-web-page).
The data reflect the average of three trials with a standard deviation of less than 2%. The table shows
6I do not report performance information for the slower network and sound cards to avoid unfairly biasing the resultsin favor of Nooks.
7The serve-complex-web-page test uses the restart recovery manager as well.
91
that Nooks achieves between 44% and 100% of the performance of native Linux for these tests.
To separate the costs of isolation and recovery, I performed performance tests on the device
drivers both with the restart recovery manager, which does no logging, and the shadow driver re-
covery manager, which logs requests to drivers. The performance and CPU utilization with the
two recovery managers was within one percent, indicating that shadow drivers and their associated
mechanisms do not add to the performance cost of using Nooks. Rather, it is the cost of Nooks’
isolation services that determines performance.
As the isolation services are primarily imposed at the point of the XPC, the rate of XPCs offers
a telling performance indicator. Thus, the benchmarks fall into three broad categories characterized
by the rate of XPCs: low frequency (a few hundred XPCs per second), moderate frequency (a few
thousand XPCs per second), and high frequency (tens of thousands of XPCs per second). I now
look at each benchmark in turn.
6.6.1 Sound Benchmark
The mp3 player application plays an MP3 file at 128 kilobits per second through the system’s sound
card, generating only 470 XPCs per second. The audio recorder application records sound at 44,100
16-bit samples per second, and generates a similarly low XPC rate. At this low rate, the addi-
tional overhead of Nooks is imperceptible, both in terms of execution time and CPU overhead. For
the many low-bandwidth devices in a system, such as keyboards, mice, Bluetooth devices [95],
modems, and sound cards, Nooks offers a clear benefit by improving driver reliability with almost
no performance cost.
6.6.2 Network Benchmarks
The network receive benchmark is an example of a moderate XPC-frequency test. Network receive
performance was measured with the netperf [116] performance tool, where the receiving node used
an isolated Ethernet driver to receive a stream of 32KB TCP messages using a 256KB buffer. The
Ethernet driver for the Intel Pro/1000 card batches incoming packets to reduce interrupt, and hence
XPC, frequency. Nevertheless, the receiver performs XPCs in the interrupt-handling code, which is
on the critical path for packet delivery. This results in a no change in throughput but an overall CPU
92
utilization increase of 7 percentage points.
In contrast, the network send benchmark (also measured using netperf) is a high XPC-frequency
test that isolates the sending node’s Ethernet driver. Unlike the network receive test, which benefits
from the batching of received packets, the OS does not batch outgoing packets. Therefore, although
the total amount of data transmitted is the same, the network send test executes almost four times
more XPCs per second than network receive test. The overall CPU utilization on the sender thus
doubles from about 28% on native Linux to 57% with Nooks. As with the network receive bench-
mark, throughput is almost unchanged. Despite the higher XPC rate, much of the XPC processing
on the sender is overlapped with the actual sending of packets, mitigating some of the overhead
from Nooks. Nevertheless, on slower processors or faster networks, it may be worthwhile to batch
outgoing packets as is done, for example, in network terminal protocols [85].
6.6.3 Database Queries Benchmark
The database benchmark is a disk-intensive application chosen to stress the disk driver. It runs
the mySQL database [122] with the Wisconsin database query benchmark [61]. Similar to the
sound benchmarks, this benchmark generates a low XPC rate, approximately 300 XPCs per second.
However, the CPU utilization during this test rose from 21% on Linux-native to 27% on Linux-
Nooks. I analyze the cause of this slowdown when discussing the compile benchmark in the next
section.
6.6.4 Compile Benchmark
As an indication of application-level file system and disk performance, I measured the time to untar
and compile the Linux kernel, both on a local ext3 file system over the ide-disk driver and on a
local VFAT file system. Table 6.5 shows that the compilation time was unchanged when ide-disk
was isolated but ran about 10% slower when VFAT was isolated. In both cases, the CPU was nearly
100% utilized in the native case, so the additional overhead resulting from Nooks is directly reflected
in the end-to-end execution time.
I take this benchmark as an opportunity to analyze the causes of Nooks’ slowdown. For ide-
disk, the low XPC rate of 58 per second shows why performance was unchanged, so I focus instead
93
on VFAT. There are two possible reasons that the compile-local benchmark runs more slowly with
Nooks: there is more code to run, and the existing code runs more slowly. To understand both
sources of slowdown, I profiled the compile-local benchmark on the native Linux kernel and Linux
with Nooks isolation.
I used statistical profiling of the kernel to break the execution time of the benchmark into com-
ponents. The results are shown in Figure 6.4. User time (not shown) for all configurations was
identical, about 477 seconds; however, kernel time is significantly different. For native Linux, the
benchmark spends a total of 39 seconds in the kernel. With Nooks, the benchmark spends more
than 111 seconds in the kernel. Most of the additional 72 seconds spent in the kernel with Nooks is
caused by executing extra code that is not executed without Nooks. The new code in Nooks (shown
by the upper bars in the Figure 6.4), accounts for 46 of the additional 72 seconds spent in the kernel
under Nooks.
Of the 46 seconds spent in Nooks, more than half (28 seconds) is due just to XPC. Object
tracking is another large cost (6 seconds), because Nooks consults a hash table to look up function
parameters in the file system interface. Other components of Nooks, such as wrappers, synchroniz-
ing page tables with the kernel, and data copying have only minor costs.
These additional execution components that come with Nooks reflect support for isolation and
recovery. The isolation costs manifest themselves in terms of XPC overhead, page table synchro-
nization, and copying data in and out of protection domains. Nooks’ support for recovery, with
either the restart or the shadow driver recovery managers, is reflected in terms of the time spent
tracking objects that occurs on every XPC in which a pointer parameter is passed. Tracking al-
lows the recovery managers to recover kernel resources in the event of an extension failure. These
measurements demonstrate that recovery has a substantial cost, or, more generally, that very fast
inter-process communication (IPC) has it limits in environments where recovery is as important as
isolation.
New code in Nooks is not the only source of slowdown. The lower bars in Figure 6.4 show
the execution time for native Linux code in the kernel (labeled other kernel) and in the VFAT file
system. For native Linux, 35 seconds were spent in the kernel proper and 4.4 seconds were spent in
the VFAT code. In contrast, with Nooks the benchmark spent about 54 seconds in the kernel and 12
seconds in VFAT. The benchmark executes the same VFAT code in both tests, and most of the kernel
94
Time spent in kernel mode for compile-local benchmark
0
30
60
90
120
Native Linux Nooks no-PT Nooks
Configuration under test
Kern
el
tim
e (
seco
nd
s)
WrappersPT Sync.Misc.Data copyingObject TrackingXPCVFATOther Kernel
Nooks
VFAT
Figure 6.4: Comparative times spent in kernel mode for the Compile-local (VFAT) benchmark.
code executing is the same as well (Nooks makes a few additional calls into the kernel to allocate
resources for the object tracker). The slowdown of VFAT and the kernel must therefore be caused
by the processor executing the same code more slowly, due to micro-architectural events such as
cache and TLB misses that Nooks causes.
I speculated that the both the high cost of XPC and the slowdown of existing kernel and VFAT
code are due to the domain change, which is very expensive on the x86 architecture. Changing the
page table register during the domain change takes many cycles to execute and causes the TLB to
be flushed, causing many subsequent TLB misses. To further understand this effect, I built a special
version of Nooks called Nooks no-PT that does not use a private page table. That is, all of the
mechanism is identical to normal Nooks, but there is no address space switch between the driver
and kernel domains. Comparing Nooks no-PT with normal Nooks thus allows us to isolate the cost
of the page table switch instruction and the subsequent TLB misses caused by the TLB flush from
other Nooks-induced slowdowns.
As shown by the center bar in Figure 6.4, the Nooks kernel without a separate page table (Nooks
no-PT) spent 51 seconds in the kernel proper and 9 seconds was in VFAT. During the run with Nooks,
the system performed 10,934,567 XPCs into the kernel and 4,0086,586 XPCs into the extension.
The cost of Nooks was reduced to just 23 seconds, and the time spent doing XPCs dropped from
95
28 seconds to 7 seconds. These results demonstrate that changing the page table is an expensive
operation, causing XPC to take many cycles, but that the TLB and cache misses caused by the
change account for only 32% of the slowdown for VFAT and 17% of the kernel slowdown. Since
the code in both configurations is the same, I speculate that the remaining performance difference
between native Linux and Nooks is caused by additional TLB and cache misses that occur because
Nooks copies data between the file system and the kernel. The Pentium 4 has just an 8KB L1 data
cache, so it is particularly sensitive to additional memory accesses.
I have done little to optimize Nooks, but believe that significant speedup is possible through
software techniques that have been demonstrated by other researchers, such as finely tuned data
structures [178, 65] and request batching.
6.6.5 Web Server Benchmarks
The final two benchmarks illustrate the impact of Nooks on server performance of transactional
workloads. Serve-simple-web-page uses a high XPC-frequency extension (kHTTPd) on the server to
deliver static content cached in memory. I used httperf [147] to generate a workload that repeatedly
requested a single kilobyte-sized Web page. kHTTPd on native Linux can serve over 15,000 pages
per second. With Nooks, it can serve about 6,000, representing a 60% decrease in throughput.
Two elements of the benchmark’s behavior conspire to produce such poor performance. First,
the kHTTPd server processor is the system bottleneck. For example, when run natively, the server’s
CPU utilization is nearly 96%. Consequently, the high XPC rate slows the server substantially.
Second, since the workload is transactional and non-buffered, the client’s request rate drops as a
function of the server’s slowdown. By comparison, the network send benchmark, which exhibits
roughly the same rate of XPCs but without saturating the CPU, degrades by only 10%. In addition,
the network send benchmark is not transactional, so network buffering helps to mask the server-side
slowdown.
Nevertheless, it is clear that kHTTPd represents a poor application of Nooks: it is CPU-limited
and performs many XPCs. This service was cast as an extension so that it could access kernel
resources directly, rather than indirectly through the standard system call interface. Since Nooks’
isolation facilities impose a penalty on those accesses, performance suffers. I believe that other
96
types of extensions, such as virus and intrusion detectors, which are placed in the kernel to access
or protect resources otherwise unavailable from user level, would make better candidates as they do
not represent system bottlenecks.
In contrast to kHTTPd, the second Web server test (Serve-complex-web-page) reflects moderate
XPC frequency. Here, we ran the SPECweb99 workload [189] against the user-mode Apache 2.0
Web Server [5], with and without Nooks isolation of the e1000 Ethernet driver. This workload
includes a mix of static and dynamic Web pages. When running without Nooks, the Web server
handled a peak of 114 requests per second. 8 With Nooks installed and the Ethernet driver isolated on
the server, peak throughput dropped by about 3%, to 110 requests per second. This small slowdown
is explained by the moderate XPC rate of almost 2000 per second.
6.6.6 Performance Summary
This section used a small set of benchmarks to quantify the performance cost of Nooks. For the de-
vice drivers, Nooks imposed a performance penalty of less than 3%. For the other kernel extensions,
the impact was higher, and reached nearly 60% for kHTTPd, an ad-hoc application extension. The
rate of XPCs is a key factor in the performance impact, as XPCs impose a high burden, due to cost
of flushing the x86 TLB in the current implementation. The performance costs of Nooks’ isolation
services depend as well on the CPU utilization imposed by the workload. If the CPU is saturated,
the additional cost can be significant, because there is no spare capacity to absorb Nooks overhead.
6.7 Summary
Overall, Nooks provides a substantial reliability improvement at low cost for common drivers. The
results demonstrate that: (1) the performance overhead of Nooks during normal operation is small
for many drivers and moderate for other extensions, (2) applications and the OS survived driver
failures that otherwise would have caused the OS, application, or both to fail. Overall, Nooks and
shadow drivers prevented 99% of system crashes and 98% of application failures, with an average
performance cost of 1% for drivers. The additional cost of shadow driver recovery as compared to
restarting a driver is negligible.
8The test configuration is throughput limited due to its single IDE disk drive.
97
These results indicate that Nooks and shadow drivers have the potential to significantly im-
prove the reliability of commodity operating systems at low cost. Given the performance of modern
systems, I believe that the benefits of isolation and recovery are well worth the costs for many
computing environments.
98
Chapter 7
THE DYNAMIC DRIVER UPDATE MECHANISM
7.1 Introduction
In addition to tolerating driver failures, Nooks and shadow drivers provide a platform for other
reliability services. The Nooks interposition service provides a mechanism to hook communication
between drivers and the kernel, as evidenced by shadow drivers’ use of it for taps. The object
tracker provides a mechanism to discover the objects in use by a driver and selectively release those
objects to the kernel. Shadow drivers provide the ability to actively proxy for drivers, concealing
their unavailability, and to transfer state between two driver instances. As a demonstration of how
these capabilities can be leveraged to provide new services, I implemented dynamic driver update,
a mechanism to replace drivers on-line on top of Nooks and shadow drivers.
As commodity operating systems become more reliable and fault-tolerant, the availability of a
system will be determined not by when it crashes, but instead by when it must be shutdown and
rebooted due to software maintenance [91, 150]. While many system components can be upgraded
on-line, critical low-level components, such as device drivers and other kernel extensions, cannot
be updated without rebooting the entire operating system. Thus, scheduled downtime promises to
be the ultimate limiter of system availability because operating systems cannot replace all running
code on-line.
Device drivers represent a large portion of this problem because they are both critical to OS
operation and frequently updated. For example, Microsoft’s Windows Update reports 115 updates
for Windows XP between 2000 and 2005, but over 480 updates for network drivers and 370 updates
for video drivers [143]. The picture is similar in Linux, where IDE storage drivers, needed by the
OS for virtual memory swapping, were updated more than 67 times in three years [196]. In addition,
studies show that patches and updates frequently cause failures, by introducing new bugs not found
in testing [192, 57].
99
To achieve high availability, operating systems must be able to replace critical device drivers
without rebooting the entire system. Furthermore, updating device drivers is inherently risky be-
cause a faulty device driver may prevent the system from booting. Thus, driver updates must be
isolated to ensure that the new code does not compromise system stability, and it must be possible
to “undo” an update that does not function.
Updating drivers on-line presents three major challenges. First, there are thousands of existing
drivers, many of which will eventually require an update. To make on-line update practical, the
system must be able to update these drivers without any work by the drivers’ authors. Existing
systems for updating code on-line require that a skilled programmer write code to transfer data
between the old and new versions [75, 66, 103, 161].
Second, because drivers assume exclusive access to devices and require access to the device
to initialize, the old driver must be disabled while the new driver initializes. Thus, the device is
unavailable during an update. Any portion of the update process that depends on a device being
present (e.g., loading the new driver off a disk when updating the disk driver) must be performed
before the old driver is disabled.
Finally, an updated driver may offer substantially different services or support different inter-
faces. For example, a new version of a driver may remove functions or capabilities provided by the
prior version. The updated device driver may not be compatible with the running version and could
cause applications or the OS to crash. Thus, the update system must ensure that an update is safe
and does not contain incompatibilities that could compromise system reliability.
In this chapter, I present the design and implementation of dynamic driver update, a service that
replaces driver code on-line. It can update common device drivers without restarting applications or
the operating system. When combined with Nooks’ isolation services, dynamic driver update also
addresses the problem that updates are not tested as completely as original code. With isolation, the
OS can tolerate any failures of the updated driver either by restarting the new version of the driver
or by rolling back to a previous version. As a result, administrators of high-availability systems can
apply important but risky updates, such as security patches, whereas today they might forgo such
patches.
In the remainder of this chapter, I present background and related work in Section 7.2. In Sec-
tion 7.3 I present the architecture and implementation of dynamic driver update, and in Section 7.4
100
I evaluate its usefulness and performance.
7.2 Related work
High-availability systems have long sought to update code on-line [75, 64]. A study of Windows
NT 4.0 installations showed that 42% of outages resulted from software maintenance. In addi-
tion, a study at IBM showed that 20% of system failures were caused by faults in poorly tested
patches [191]. Hence, the reliable upgrade of system components is critical to high availability. Pre-
vious approaches to this problem differ in the programmer’s responsibility, the update mechanisms
and the granularity of updates. I discuss each of these in turn.
Previous on-line update systems require a programmer to provide code to facilitate the update.
The programmer’s responsibilities range from following design-time rules, such as inheriting from
a specific base class, to writing functions that translate data formats when they change. The systems
that rely on a compiler to provide the mechanism for updating code require the least programmer
effort. OPUS, for example, uses a compiler to generate patches automatically from two versions of
the source code [2], but is restricted to patches that do not change data formats. Dymos [129] and
Dynamic Software Updating [103] can update arbitrary code but require that programmers write a
conversion function if data types change.
Other approaches place more responsibility on the programmer. Systems that rely on object-
oriented language features, such as indirect function calls, require that updatable code inherit from
an “updatable” virtual base class. Programming-language specific systems, such as dynamic C++
classes [105], and dynamic update for Java [161] take this approach, as does the k42 operating
system. Another task placed on programmers is to ensure that data structures remain consistent
during an update by identifying locations in the code where updates can occur safely [98]. At the
most extreme, the IROS platform includes a method to update driver code in the driver interface,
effectively pushing all the work onto the programmer [7].
In some cases, a system can reduce the programmer’s role by leveraging the existing properties
of the updatable code. For example, the Simplex system updates real-time event handlers with
no additional code by running old and new handlers in parallel until their outputs converge [84].
Microsoft’s most recent web server, IIS 7.0, relies on the independence of separate web requests to
101
replace extensions at run-time. As soon as an updated extension is available, the web server routes
requests to the new version [148].
There have been four major mechanisms used to replace updated code. First, within a cluster
of replicated servers, an administrator can restart an entire machine to load the new code while the
remaining machines provide service. The Visa payment processing system, for example, applies
20,000 updates a year with this mechanism yet achieves 99.5% availability [166]. Recent work on
virtual machines may enable this technique to be applied on a single computer [137]. The benefit
of this approach is that it is very general; anything can be updated by rebooting the whole machine.
However, it requires the most work from application programmers, as these systems rely on the
application to perform failover and state replication [93].
Second, the system can launch a copy of a running process with the new code on the same
hardware [94], or on additional hardware [181], and then copy in the state of the existing process.
This approach is limited to user-mode processes, as without the support of a virtual machine monitor,
two operating system kernels cannot coexist on a single machine.
Third, the system can patch running code by dynamically dispatching callers to functions that
have been updated. In this case, wrapper functions [75, 89, 188], dynamically dispatched func-
tions [105, 161, 66], and compilers that produce only indirect function calls [129, 103] can all
redirect callers to the new code. This approach, while requiring more infrastructure, has the benefit
of being working in many environments and with many program changes.
Finally, the system can directly modify code in memory by replacing the target address of a
function call [2] or by replacing any instruction with a branch to new code [120]. This approach is
limited to code changes and cannot handle data format changes.
Several different approaches have been used to dynamically update operating system code. The
k42 microkernel operating system is built upon object-oriented components that can be hot-swapped
and updated dynamically [188, 21]. The k42 update system relies on the unique object model of
k42 and hence is difficult to apply to commodity operating systems that are not object-oriented [21].
The IROS real-time operating system includes a method in the driver interface that allows two
drivers to coexist simultaneously and to transfer state [7]. The operating system acts as a wrapper
and redirects calls to the new driver. Thus, driver writers must provide code to support updating
drivers. Kishan and Lam applied instruction editing to the Mach kernel for both extensibility and
102
optimization [120]. As previously mentioned, this approach only can only apply changes that do
not modify data formats and hence is not broadly applicable.
Dynamic driver update provides on-line update for device drivers. As with Nooks, my paramount
goal is compatibility with existing code. Dynamic driver update relies on shadow drivers to update
a device driver in-place without assistance from a programmer. My update system uses the com-
mon interfaces of device drivers to capture and transfer their state between versions automatically.
Thus, compatibility reduces the deployment cost by removing the programmer support requirement
of previous approaches.
7.3 Design and Implementation
In keeping with the Nooks design principles (see Section 4.1), dynamic driver update (DDU) reflects
a practical, systems-oriented approach to updating device drivers on-line. Existing approaches to
dynamically updating code rely on the compiler, programming language, or programmer to perform
the update. In contrast, dynamic driver update is a service that updates drivers on their behalf.
The design of DDU shares the principles of shadow drivers:
1. Update logic should be generic.
2. Device driver updates should be concealed from the driver’s clients.
3. Update logic should be centralized.
4. Update services should have low overhead when not needed.
and adds one more:
5. Safety first. The system may disallow an update if it is not compatible with running code.
The update system must work with existing drivers because it is too hard, expensive, and error-prone
to add on-line update support to drivers. With a systems-oriented approach, a single implementation
can be deployed and used immediately with a large number of existing drivers. Thus, I centralize
the update code in a single system service. As a common facility, the update code can be thoroughly
examined and tested for faults. In contrast, practical experience has shown us that many drivers do
not receive such scrutiny, and hence are more likely to fail during an update [57].
103
These principles limit the applicability of dynamic driver update. My solution applies only to
drivers that can load and unload dynamically. Furthermore, my design only applies to drivers that
share a common calling interface. A driver that uses a proprietary or ad-hoc interface cannot be
updated automatically without additional code specialized to that driver. Finally, if the new driver
supports dramatically different capabilities than the existing driver, the update will fail because the
changes cannot be hidden from the OS and applications.
Because of these limitations, I consider dynamic driver update to be an optimization and preserve
the possibility of falling back to a whole-system reboot if necessary.
7.3.1 Design Overview
Dynamic driver update replaces a driver by monitoring the state of a driver as it executes, loading
the new code into memory, starting the new driver, and transferring the old driver’s state into the
new driver. During the update, the system impersonates the driver, ensuring that its unavailability
is hidden from applications and the OS. As the new driver starts, the update system connects the
new driver to the old driver’s resources and verifies that the new driver is compatible with the old
driver. If it is not compatible, the system can roll back the update. In essence, dynamic driver update
replaces the driver code and then reboots the driver.
Similar to recovering from driver failures, dynamic driver update seeks to conceal the update
from the OS and applications. Unlike a whole-system reboot, where the ephemeral state of a driver
is lost, restarting just the driver requires that this state be preserved so that application and OS re-
quests can be processed correctly. Hence, dynamic driver update leverages shadow drivers to restart
an updated device driver without impacting applications or the OS. Because it is not concerned with
fault tolerance, dynamic driver update uses only the object tracking and interposition services nec-
essary for shadow drivers. The object tracker records kernel objects in use by the driver for later
reclamation during an update. The interposition service supports taps, which invoke the shadow on
every call between a driver and the kernel. Isolation may optionally be used when the updated code
is potentially faulty. Dynamic driver update also leverages shadow drivers to obtain the state of a
driver and to transfer state between driver versions.
Table 7.1 lists the steps of updating a driver. The process begins when a driver is initially
104
Table 7.1: The dynamic driver update process.
Update Steps1. Capture driver state2. Load new code3. Impersonate driver4. Disable/garbage collect old driver5. Initialize new driver6. Transfer in driver state7. Enable new driver
loaded and the system starts capturing its state and ends after when the new driver begins processing
requests. I now describe each of the update steps in more detail.
7.3.2 Capturing Driver State
From the moment a driver is first loaded, a shadow driver monitors its state. The state of a device
driver consists of: (1) the static global variables and constants in the driver code, (2) the state
it receives from hardware, and (3) the state it receives from the kernel. The state from code is
static, and hence does not need to be preserved across an update. From hardware, a driver receives
persistent data and environmental parameters, such as the speed of a network link. For most drivers
either this information is stable (i.e., does not change over time) or the driver maintains no history of
this information and reacquires it as needed. As a result, this state need not be explicitly preserved
when updating a driver.
From the kernel, the driver receives configuration and I/O requests. While configuration changes
must persist across an update, completed I/O requests generally have no impact on future requests.
Hence, they need not be retained after a driver update. Thus, the only driver state that must be
preserved across an update is the history of configuration requests and the requests it is currently
processing.
Dynamic driver update depends on shadow drivers and the object tracker in Nooks to capture
the state of a device driver. The object tracker records kernel resources in use by a driver, so that
the update system can disable and garbage collect the old driver. The shadow logs configuration
requests and in-progress I/O requests. With this information, the shadow can restore the driver’s
105
Disk Device Driver
Version 1
Shadow Disk
Driver
OS Kernel
Disk Drive
Disk Device Driver
Version 2
Figure 7.1: A shadow driver capturing the old driver’s state.
internal state after an update. The class-based nature of device drivers allows a single shadow driver
to capture the state of any driver in its class. As a result, no code specific to a single driver is required
for capturing driver state. Figure 7.1 shows a shadow driver capturing driver state by monitoring
communication between the kernel and the driver. In this figure, a new driver version has been
pre-loaded into memory for an upcoming update.
7.3.3 Loading Updated Drivers
A system operator begins an update by loading the new driver code. As discussed in Section 7.1, the
services of a driver are not available while it is restarting. This is an issue, for example, if the update
mechanism relies on disk access after the disk driver has been disabled. To avoid such circular
dependencies, an operator must pre-load the new driver version into memory so that the system can
update drivers that store their own updates.
Dynamic driver update logically replaces the entire driver at once. While the actual change
to a driver may be localized to a single function, replacing the entire driver allows the system to
leverage existing OS support for loading and initializing drivers. When updating only a subset of
the modules comprising a multi-module driver, the update system optimizes the in-memory footprint
by replacing only those modules whose memory images change. In addition to new modules, this
includes any module that links to the new module, because function addresses may have changed.
106
Figure 7.1 shows the new version loaded in memory alongside the existing version. Once the
new code is in memory, an operator notifies the update system to replace the driver. The driver
update system leaves the old driver code in memory until the update completes, which allows the
system to revert back to the old driver if the update fails. For example if the new driver does not
initialize or is not compatible with the old driver, the update system restarts the old driver.
7.3.4 Impersonating the Driver
Dynamic driver update relies on shadow drivers to impersonate the old driver during the update. I
previously described the process in Section 5.3.4. The shadow driver actively proxies for the driver,
ensuring that applications and the OS do not observe that it is unavailable. For systems that update
user-mode code, the application can be suspended during the update process. However, drivers are
frequently called at interrupt level, where calls cannot be suspended, and hence active proxying is
necessary.
7.3.5 Garbage Collecting the Old Driver
The goal of this stage is to release any resources in use by the old driver that are not needed as
part of the update process. As with impersonating the driver, shadow drivers perform this function.
Those objects that the kernel relies on to invoke the driver, such as driver registration objects, and
the hardware resources in use by a driver, such as interrupt-request lines and I/O memory regions,
are retained for restarting the driver. The shadow releases everything else, including dynamically
allocated memory and other support objects, such as timers. Section 5.3.3 contains the details. After
this stage, both old and new driver are present in memory in a clean, uninitialized state and either
one can be started. Starting the old driver is equivalent to recovering after a failure, while starting
the new one updates the driver code.
7.3.6 Initializing the New Driver
Once the old driver is garbage collected, the shadow driver initializes the new driver with the same
sequence of calls that the kernel makes when initializing a driver. Thus, the new driver starts exactly
as if the kernel initialized it. For example, when starting a network device driver, the shadow calls the
107
driver’s init module function, followed by the open function to initialize the network, followed
by set multicast list. In addition, the shadow also calls internal kernel routines to ensure the
network driver is properly bound into the kernel, such as dev activate and netif schedule
to notify the kernel that the driver is prepared to start sending packets. The sequence of calls is
hardwired into the shadow driver for a particular class of drivers. Hence, if the kernel evolves and
makes additional calls during driver initialization, the shadow must similarly be updated. During
this initialization stage, taps send all calls made by the updated device driver to the shadow instead
of the kernel, allowing the shadow to connect it to the resources of the old driver.
The benefit of replicating the kernel’s initialization sequence is that no new code is needed in the
driver for an update. Instead, dynamic driver update repurposes the driver’s existing initialization
code by controlling the driver’s environment during restart, spoofing it into believing that the kernel
is starting it normally.
The update system faces two major challenges in starting the new driver: identifying existing
kernel objects that the new driver requests, and ensuring that the new driver is compatible with the
old driver. I now discuss each of these challenges in detail.
Updating References
When the new driver requests a resource from the kernel or provides a resource to the kernel, the
shadow must locate the corresponding resource belonging to the old driver. The shadow calls into
the update system to perform this mapping. Once located, the update system updates the resource
to reference the new driver. The primary challenge in this process is identifying which resource of
the old driver corresponds to the resource used by the new driver.
Drivers usually request kernel resources with a unique identifier, which allows the update system
to locate the old driver’s corresponding resource. For example, drivers provide a unique name,
device number, or MAC address when registering with the kernel, and unique memory addresses
when requesting I/O regions. The update system uses the unique identifier and resource type to
locate the old driver’s resource in the object tracker.
In some cases, the driver passes a resource to the kernel without unique identification. The
update system then uses the context of the call (i.e., the stage of the restart) to make the association.
108
For example, when the kernel opens a connection to a sound-card driver, the driver returns a table of
function pointers with no identifying device name or number. To transfer these function tables to the
new driver, the shadow calls into the new driver to re-open every connection. The new driver returns
its function table. Because the shadow driver knows that it is re-opening a connection, it notifies the
update system to replace the function table in the existing connection object with the new functions
provided by the driver. This contextual approach to matching resources can be applied elsewhere,
for example, when updating a network driver that lacks a unique MAC address.
After locating the appropriate kernel resource, the update system modifies it to reflect differences
between the old and new driver. For example, the new driver may provide additional functions or
capabilities. In some cases, the update system need not update the kernel object directly. Instead, it
relies on the layer of indirection provided by wrappers and the object tracker to update the mapping
between kernel and driver versions. For example, when updating interface function pointers, the
update system replaces the mapping maintained by the object tracker rather than changing function
pointers in the kernel. If this is not possible, the update system locks and updates the kernel object
directly.
Detecting Incompatibilities
A second issue that arises when updating code without programmer support is ensuring that the new
code is compatible with the running code. If an update is not compatible, it may cause the OS and
applications to fail. Incompatibilities arise when a new driver implements different functions (e.g.,
higher-performance methods), different interfaces (e.g., a new management interface), or different
options for an existing function. Complicating matters, applications may expose incompatibilities
long after a new driver starts. For example, an application may use capabilities it queried from the
old driver that are not present in the new driver. Reflecting the “safety first” principle, the update
system must detect whether the update could cause callers to function incorrectly in order to prevent
applications from seeing unexpected behavior.
Drivers express their capabilities in two ways. First, drivers may provide only a subset of the
functions in an interface. For example, a driver writer may choose not to implement certain high-
performance methods. Second, drivers can express their capabilities through feature fields. For
109
example, network drivers pass a bit-mask of features, such as the ability to offload TCP checksum-
ming and do scatter/gather I/O, to the kernel. The update system considers an update compatible
only if a new driver provides at least the same capabilities as the old driver.
As part of transferring references to the new driver, the update system performs three compati-
bility checks. First, it verifies that the new driver provides at least the same interfaces to the kernel
as the old driver. If a driver has not registered all of the old driver’s interfaces during initialization, it
is incompatible. Second, the update system verifies that, within an interface, the new driver imple-
ments at least the functions that the old driver provided. If so, the system updates the object tracker
to reference the new driver. Third, for drivers with features that can be queried by applications,
the update system explicitly queries the new driver’s feature set during an update and compares it
against the old driver’s feature set. Thus, the update system detects incompatibilities during the
update process, ensuring that there will not be latent problems discovered long after the process
completes.
When it detects an incompatible update, the update system has three options. It can take a
conservative approach and revert to the existing driver. Or, it can let the update go forward but fail
any future request that would reveal the incompatibility. The final option is to forgo compatibility
checks and allow applications to observe the incompatibilities. These options allow an operator to
prioritize an important update (for example, one that prevents unauthorized use of the system) over
compatibility. Lower-priority incompatible updates can still be applied with a whole-system reboot.
Rolling back an update presents compatibility problems as well. If the new driver provides
a superset of the old driver’s capabilities, it may not be possible to roll back an update without
notifying applications that observed the new driver’s capabilities. To preserve the option to roll
back, the update system can optionally conceal the new capabilities of a driver by masking out
new capabilities or functions to equal those in the old driver. Once the update has been tested, an
administrator can apply it again without concealing its capabilities. This ensures that a potentially
faulty update can be tested without risking system stability.
For some driver updates, these rules are overly conservative. For example, if a driver provides
a function or capability that is never used, a new version need not provide that function. However,
the update system cannot know whether an application will call the function later, and hence uses
conservative checks.
110
Disk Device Driver
Version 1
Shadow Disk
Driver
OS Kernel
Disk Drive
Disk Device Driver
Version 2
Figure 7.2: Device driver and shadow after an update.
7.3.7 Transferring Driver State
While previous on-line update systems require applications to transfer their own state between ver-
sions, dynamic driver update transfers driver state automatically, without the involvement of driver
writers. As discussed in Section 5.3.3, the shadow driver captures the state of a driver by logging
its communication with the kernel. It then transfers the old driver’s state into the new driver by
replaying the log, using the same calls into the driver that it previously monitored. Shadow drivers’
ability to automatically capture and restore the state of unmodified drivers is critical to providing
dynamic update without programmer involvement.
7.3.8 Enabling the New Driver
The final step of an update is to enable the new driver. First, the shadow submits any requests that it
queued during the update, and then it notifies the taps that requests should be sent to the real driver
instead of the shadow driver. After the new driver is running, operators may either free the memory
occupied by the old driver code or leave it in memory if they desire the opportunity to roll back the
update later. The system state after an update is shown in Figure 7.2. In this figure, the kernel calls
into the new driver while the old driver remains in memory in case the update must be rolled back.
111
Table 7.2: The number of lines of code dynamic driver update.
Source Components # LinesLinux kernel 181Nooks 17Shadow drivers 51Dynamic driver update 402
Total number of lines of code 651
7.3.9 Implementation Size
I constructed dynamic driver update as a service on top of Nooks and shadow drivers. As result,
little additional code was necessary to implement this service. Table 7.2 shows the breakdown
of code between changes to the kernel, changes to Nooks isolation code, changes to shadow driver
recovery code, and new code components. The kernel had to be changed to support loading multiple
versions of a driver simultaneously. The changes to Nooks allow waiting for the driver to become
quiescent instead of preempting threads currently executing in the driver. The most substantial
changes were to the shadow driver code, in which I added new call-outs for mapping between old
and new versions of driver resources and for compatibility checks. The small size of the entire
implementation demonstrates the usefulness of Nooks and shadow drivers as a platform, as without
this code the entire service would have been many times larger.
7.3.10 Summary
Dynamic driver update takes a systems-oriented approach to on-line update by providing a cen-
tralized service to update unmodified drivers. Before an update, it loads driver code into memory
to avoid circular dependencies. During an update, a shadow driver hides the driver’s unavailabil-
ity from applications and the OS by handling requests on the driver’s behalf. After an update, the
system presents the option of rolling back to the old version automatically if the new version fails.
Dynamic driver update relies on existing driver interfaces to update the driver and transfer the
state of the current driver. As the new driver initializes, the update system connects it to the resources
of the old driver. The update system reassigns kernel structures referencing the old driver to the new
driver. Once the new driver has initialized, the shadow driver transfers in the state of the old driver
112
by replaying its log of requests and by re-opening connections to the driver. Finally, the update
system dynamically detects whether the old and new drivers are compatible by comparing their
capabilities. If the new driver is not compatible, either the update or the action that would reveal
the incompatibility can be canceled. When used with Nooks isolation services, the system allows
potentially faulty updates to be tested without compromising availability.
Dynamic driver update is built on top of Nooks and shadow drivers. Little new code was needed
to implement this service, as Nooks and shadow drivers provide object tracking, interposition, active
proxying, and restarting services. Thus, dynamic driver update demonstrates that Nooks and shadow
drivers are a useful platform for implementing additional reliability and availability services.
7.4 Evaluation of Dynamic Driver Update
In this section I answer three questions about dynamic driver update:
1. Transparency. Can dynamic driver update replace existing drivers?
2. Compatibility. Can it detect and roll back incompatible updates?
3. Performance. What is the performance overhead of DDU during normal operation (i.e., in
the absence of updates), and are on-line updates significantly faster than rebooting the whole
system?
Based on a set of experiments using existing, unmodified device drivers and applications, my results
show that dynamic driver update (1) can apply most released driver updates, (2) can detect when an
update is incompatible with the running code and prevent application failures by rolling back the
update, and (3) imposes only a minimal performance overhead.
7.4.1 Methodology and Platforms
I implemented dynamic driver update in the Linux 2.4.18 operating system kernel. I produced two
OS configurations based on the Linux 2.4.18 kernel:
1. Linux-Native is the unmodified Linux kernel.
2. Linux-DDU is a version of Linux that includes the Nooks wrappers, object tracker, shadow
drivers plus the dynamic driver update code. It does not include the Nooks isolation services.
113
Table 7.3: The three classes of shadow drivers and the Linux drivers tested.
Class Driver DeviceSound emu10k1 SoundBlaster Live! sound card
audigy SoundBlaster Audigy sound cardes1371 Ensoniq sound card
Network e1000 Intel Pro/1000 Gigabit Ethernet3c59x 3COM 3c509b 10/100 Ethernetpcnet32 AMD PCnet32 10/100 Ethernet
Storage ide-disk IDE disk
I ran the experiments on a 3 GHz Pentium 4 PC with 1 GB of RAM and a single 80 GB, 7200 RPM
IDE disk drive. Networking tests used two identically configured machines.
I tested dynamic driver update with the same three shadow drivers used previously in Chapter 6:
sound card, network interface controller, and IDE storage device. To ensure that dynamic driver
update worked consistently across device driver implementations, I tested it on seven different Linux
drivers, shown in Table 7.3. Although I present detailed results for only one driver in each class
(shown in boldface, emu10k1, e1000, and ide-disk), behavior across all drivers was similar.
Of these three classes of drivers, only IDE storage drivers require a full system reboot in Linux
to be updated. However, updating members of the other two classes, sound and network, disrupts
applications. Dynamic driver update improves the availability of applications using these drivers,
because they need not be restarted when a driver is updated.
7.4.2 Effectiveness
To be effective, the dynamic driver update system must hide updates from the applications and the
OS, apply to most driver updates, and detect and handle incompatible updates. Dynamic driver up-
date depends on shadow drivers to conceal updates. As a result, it shares their concealment abilities.
Section 6.4 demonstrated that shadow drivers could conceal recovery from all tested applications in
a variety of scenarios. As a result, I only examine the latter two requirements: whether dynamic
driver update can apply existing driver updates and whether it can detect and roll back incompatible
updates.
114
Table 7.4: Tested drivers, the number of updates, and the results.
Driver Module # of updates # Failed
emu10k1 emu10k1 10 0ac97 codec 14 0soundcore 6 0
e1000 e1000 10 1ide-disk ide-disk 17 0
ide-mod 17 0ide-probe-mod 17 0
.
Applicability to Existing Drivers
I tested whether dynamic driver update can update existing drivers by creating a set of different
versions of the drivers. I compiled past versions of each driver, which I obtained either from the
Linux kernel source code or from vendor web sites. For drivers that consist of multiple modules,
I created separate updates for each module and replaced the modules independently. Because the
driver interface in the Linux kernel changes frequently, I selected only those drivers that compile
and run correctly on the 2.4.18 Linux kernel. Table 7.4 show the modules comprising each of my
tested drivers, the number of updates for each module, and the number of updates that failed due to
compatibility checks.
Using the Linux-DDU kernel, I applied each driver update from oldest to newest while an ap-
plication used the driver. For the sound-card driver, I played an mp3 file. For the network drivers, I
performed a network file copy, and for the IDE storage driver I compiled a large program. For each
update, I recorded whether the DDU successfully updated the driver and whether the application
and OS continued to run correctly. If the update process, the application or the OS failed, I consider
the update a failure.
The results, shown in Table 7.4, demonstrate that dynamic driver update successfully applied 90
out of 91 updates. In one case, when updating the e1000 driver from version 4.1.7 to 4.2.17, the
shadow could not locate the driver’s I/O memory region when the new driver initialized. The new
driver used a different kernel API to request the region, which prevented the update system from
finding the region in the object tracker. As a result, the new driver failed to initialize and caused the
update system to roll back. The file-copy application, however, continued to run correctly using the
115
old driver.
This single failure of dynamic driver update demonstrates the complexity of transferring kernel
resources between driver versions. Shadow drivers require a unique identifier to locate the resource
of an old driver that corresponds to a new driver’s request. Some of the unique identifiers I chose
depend on the API used to request the resource. When a driver changes APIs, the shadow cannot
locate the old driver’s resources during an update. Ideally, the identifier should be independent of
the API used to request the resource. In practice, this is difficult to achieve within the existing Linux
kernel API, where there are many resources that can be accessed through several different APIs.
Compatibility Detection
While I found no compatibility problems when applying existing driver updates, compatibility de-
tection is nonetheless critical for ensuring system stability. To test whether dynamic driver update
can detect an incompatible update, I created a new version of each of the tested drivers that lacks a
capability of the original driver. I created emu10k1-test, a version of the emu10k1 driver that sup-
ports fewer audio formats. The update system detects the sound-card driver’s supported formats by
querying the driver’s capabilities with an ioctl call after it initializes. From the e1000 driver, I
created e1000-test that removes the ability to do TCP checksumming in hardware. The driver passes
a bit-mask of features to the kernel when it registers, so this change should be detected during the
update. Finally, I created ide-disk-test, which removes the ioctl function from ide-disk . This
change should be detected when the driver registers and provides the kernel with a table of function
pointers.
For each test driver, I first evaluate whether compatibility detection is necessary by updating
to the test driver with the compatibility checks disabled. Similarly to the previous tests, I ran an
application that used the driver at the time of update. With compatibility checking enabled, I then
applied the same update. For each test, I recorded whether the system applies the update and whether
the application continued to run correctly.
Table 7.5 shows the outcome of this experiment. With compatibility checking disabled, all three
drivers updated successfully. The sound application failed, though, because it used a capability of
the old driver that the new driver did not support. The disk and network applications continued to
116
Table 7.5: The results of compatibility tests.
Driver No Compatibility Check? Compatibility Check?emu10k1-test Application fails Update rolled backe1000-test Works Update rolled backide-disk-test Works Update rolled back
execute correctly because the kernel checks these drivers’ capabilities on every call, and therefore
handles the change in drivers gracefully. With compatibility checking enabled, dynamic driver up-
date detected all three driver updates as incompatible and rolled back the updates. The applications
in these tests continued to run correctly.
These tests demonstrate that compatibility checking is important for drivers whose capabilities
are exposed to applications. The sound application checks the driver’s capabilities once, when
starting, and then assumes the capabilities do not change. In contrast, only the kernel is aware of the
network and disk drivers’ capabilities. Unlike the sound application, the kernel checks the driver
capabilities before every invocation of the driver, and hence the compatibility check during update
is not necessary.
7.4.3 Performance
Relative to normal operation, driver updates are a rare event. As a result, an on-line update system
must not degrade performance during the intervals between updates. In addition, updates must occur
quickly to achieve high availability. Dynamic driver update introduces overhead from taps, which
invoke the shadow on every function call between the kernel and drivers, and from the shadow driver
itself, which must track and log driver information.
I selected a variety of common applications that use the three device driver classes. The applica-
tion names and workloads are shown in Table 7.6. I measured their performance on the Linux-Native
and Linux-DDU kernels. This test measures the cost of the update system alone, without isolation.
If the updated driver is potentially faulty, then Nooks isolation should be used as well. Because dy-
namic driver update adds no additional overhead beyond the logging performed by shadow drivers,
the performance with isolation is the same as reported in Section 6.6.
117
Table 7.6: The applications used for evaluating dynamic driver update performance.
Device Driver Application ActivitySound • mp3 player (zinf) playing 128kb/s audio
(emu10k1 driver) • audio recorder (audacity) recording frommicrophone
Network • network send (netperf) over TCP/IP(e1000 driver) • network receive (netperf) over TCP/IP
Storage • compiler (make/gcc) compiling 788 C files• database (mySQL) processing the Wisconsin
Benchmark
Different applications have different performance metrics of interest. For the sound-card driver,
I measured the available CPU while the programs ran. For the network driver, throughput is a more
useful metric; therefore, I ran the throughput-oriented network send and network receive bench-
marks. For the IDE storage driver, I measured elapsed time to compile a program and to perform a
standard database query benchmark. I repeated all measurements several times and they showed a
variation of less than one percent.
Figure 7.3 shows the performance of Linux-DDU relative to Linux-Native, and Figure 7.4 shows
the CPU utilization of the tests on both platforms. These figures make clear that dynamic driver
update imposes only a small performance penalty compared to running on Linux-Native. Across all
six applications, performance on Linux-DDU averaged 99% of the performance on Linux-Native.
However, CPU utilization varied more, rising by nine percentage points for the network send bench-
mark. As a comparison, the average CPU utilization increase with Nooks isolation was 5%, while
with dynamic driver update it was just 2%. Overall, these results show that dynamic driver update
imposes little run-time overhead for many drivers.
As with the cost of Nooks, this overhead can be explained in terms of frequency of communi-
cation between the driver and the kernel. On each kernel-driver call, the shadow driver may have to
log a request or track an object. For example, the kernel calls the driver approximately 1000 times
per second when running audio recorder, each of which causes the shadow to update its log. For the
most disk-intensive of the IDE storage applications, the database benchmark, the kernel and driver
interact only 290 times per second. The network send benchmark, in contrast, transmits 45,000
118
Relative Performance
80
85
90
95
100
mp3 player audiorecorder
networksend
networkreceive
compiler database
Rela
tive P
erf
orm
ance (
%)
Linux-Native Linux-DDU
Sound Network Storage
Figure 7.3: Comparative application performance on Linux-DDU relative to Linux-Native. TheX-axis crosses at 80%.
packets per second, causing 45,000 packets to be tracked. While throughput decreases only slightly,
the additional work increases the CPU overhead per packet by 34%, from on 3µs on Linux-Native
to 4µs on Linux-DDU.
Another important aspect of performance is the duration of time when the driver is not available
during an update. For each of the three drivers, I measured the delay from when the shadow starts
impersonating the device driver until it enables the new driver. The delay for the e1000 and ide-disk
drivers was approximately 5 seconds, almost all due to the time the driver spends probing hardware1.
The emu10k1 updates much faster and is unavailable for only one-tenth of a second. In contrast,
rebooting the entire system can take minutes when including the time to restart applications.
1The e1000 driver is particularly slow at recovery. The other network drivers I tested restarted in less than a second.
119
CPU Utilization
0
20
40
60
80
100
mp3 p
layer
audio
record
er
netw
ork
send
netw
ork
receiv
e
com
pile
r
data
base
CP
U U
tiliz
ation (
%)
Linux-Native Linux-DDU
Sound Network Storage
Figure 7.4: Absolute CPU utilization on Linux-DDU and Linux-Native.
7.5 Summary
While operating systems themselves are becoming more reliable, their availability will ultimately
be limited by scheduled maintenance required to update system software. This chapter presented a
system for updating device drivers on-line that allows critical system drivers, such as storage and
network drivers, to be replaced without impacting the operating system and applications, or more
importantly, availability.
Dynamic driver update demonstrates that shadow drivers and Nooks can be a platform for addi-
tional reliability services. The update system builds on this platform to update driver code in-place
with no changes to the driver itself. The system loads the new driver code, reboots the driver, and
transfers kernel references from the old driver to the new driver. To ensure that applying an update
will not reduce reliability due to changes in the driver, dynamic driver update checks the new driver
for compatibility, and can rollback incompatible updates. In testing, I found that dynamic driver
update could apply 99% of the existing driver updates and had minimal impact on performance.
120
Thus, this system reduces the downtime caused by driver maintenance.
121
Chapter 8
LESSONS
8.1 Introduction
In this chapter, I present lessons from the process of designing and implementing Nooks. While I
built Nooks within Linux, I also advised several undergraduate students on an effort to implement
the Nooks architecture within Windows 2000. The Windows implementation was not completed,
but the effort highlighted differences between the Linux and Widows kernels and provided valuable
insight into how an operating system’s structure affects its reliability.
To execute device drivers reliably, Nooks is involved on every event that impacts drivers. These
events may be explicit, such as function calls between a driver and the kernel, or implicit, such as
changes to the kernel address space. The ease of implementing a reliability layer depends largely
on the ease of identifying these events and injecting code when they occur. Thus, OS structure has
an impact on the ease of tolerating driver failures. At a high level, when relevant events are more
frequent, the performance cost of fault tolerance rises. Similarly, when the difficulty of detecting or
hooking events rises, the complexity of providing fault tolerances rises.
The Linux and Windows operating systems differ greatly in their internal organization and struc-
ture. Many of these differences have a large impact on the ease of supporting fault-tolerant execution
of drivers. The differences between the two OSs thus provide a lens for examining these structural
impacts. In this chapter, I discuss the lessons learned about the relationship between OS structure
and reliability from the two Nooks implementation efforts.
8.2 Driver Interface
The interface between the kernel and drivers affects the ease of isolating drivers and recovering from
their failure. The impact comes from the size, complexity and organization of the kernel-driver
calling interface. Smaller interfaces, with fewer functions, less-complex parameters, and explicit
122
communication reduce the coding effort to isolate drivers.
At the most basic level, every function in the driver interface requires a wrapper and object
tracking code. Thus, the amount of code is proportional to the size and complexity of the driver
interface. For example, the NDIS interface for Windows network drivers contains 147 functions
that drivers may implement; it therefore requires more coding effort than the network interface in
Linux, which defines just 21 functions. However, other factors beyond interface size determine the
total complexity of implementing this code.
Nooks must interpose on all communication between drivers and the kernel to contain driver
failures. Interposition is straightforward when every act of communication is explicit, through func-
tion calls and their parameters. The code to synchronize objects across protection domains then can
be confined to parameters passed through the kernel-driver interface. In contrast, communication
through shared memory requires additional code to detect communication channels and to detect
when communication occurs. For example, Linux provides a function for accessing the memory of
a user-mode process. Nooks can interpose on the function and provide safe access to the user-mode
data by XPCing into the kernel, which maps the calling process’ address space. In contrast, Win-
dows drivers directly access user-mode pointers. Nooks must detect this access as a communication
channel and not as a bad-pointer access. Thus, this access requires that Nooks map the user-mode
address space of client processes into driver protection domains, which is a more complex and ex-
pensive approach than the XPC-based solution on Linux.
Isolation is also simplified when the kernel-driver interface relies on abstract data types (ADTs),
because all access to an ADT goes through accessor functions. These functions provide two ben-
efits. First, they are an explicit signal that the kernel and driver are communicating, providing an
opportunity to interpose on that communication. Second, accessor functions allow the choice of
performing the operation locally or remotely in a different protection domain. For infrequently
modified objects, it is often faster to modify shared objects in the context of the kernel instead of
creating a local copy of an object, while for frequently modified objects a copy may be faster.
As previously discussed, opaque pointers in the kernel-driver interface complicate fault-tolerance
because they may point to a complex data structure not understood by wrappers and shadow drivers.
While isolation may be possible, without knowledge of the semantics of opaque parameters, trans-
parent recovery is not. For example, a shadow driver cannot know whether a request is idempotent.
123
The frequency of opaque data types varies widely between driver classes and operating systems. In
Linux, most drivers that provide a common service, such as block storage or network access, imple-
ment a common interface with no opaque or driver-specific parameters. Instead, drivers export their
additional capabilities through the /proc file system interface, which does not use opaque point-
ers. In contrast, Windows drivers provide extended capabilities through the ioctl and fsctl
interfaces, which use opaque pointers.
Finally, the complexity of the driver interface itself is a major source of bugs. If driver writers
must concern themselves with low-level details of synchronization and memory management, then
bugs are more likely. If, instead, the interface provides high-level abstractions that hide this com-
plexity, then driver code is simpler and less error-prone. For example, the Windows Driver Model
requires complex and error-prone synchronization logic for canceling a pending I/O request [145].
The driver must remove the I/O request from a queue while ensuring that it is not completed by
another thread. Unfortunately, this complex interface, provided to support high-performance I/O,
often leads to driver bugs that reduce reliability [13]. In contrast, the Linux driver interface has a
much simpler design with many fewer functions. Furthermore, the kernel performs much of locking
before invoking device drivers, so that drivers need not synchronize access to shared data structures.
Rather than require driver writers to implement complex cancellation logic, Linux relies on the Unix
signal mechanism to cancel an I/O request. Thus, simpler interfaces, while potentially reducing per-
formance for high-bandwidth and low-latency devices, may substantially improve the reliability of
the system.
8.3 Driver Organization
The organization of drivers in the kernel and their communication patterns also affect the ease of tol-
erating their failure. When drivers only communicate with the kernel, then the complete semantics
of the driver interface must be understandable to the kernel. Hence, these semantics can be used to
produce wrapper, object tracking, and recovery code. In contrast, if drivers communicate with other
drivers, additional code is needed for every driver-to-driver communication channel, which reduces
the leverage provided by the class structure of drivers.
Linux and Windows differ in their organization of drivers, and this impacts the ease of isolating
124
individual drivers. Linux uses a library model for drivers, in which drivers link with a support library
containing common routines and then export a single interface to the kernel. There is no support
for extending the driver capabilities with additional code, so there is no need to isolate cooperating
drivers from each other. In the Linux model, the driver and the library are isolated together in a
single protection domain.
In contrast, Windows employs a stacked driver model, in which an I/O request may pass through
a sequence of drivers before delivery to a device. Each driver in the layer may provide additional
service, such as RAID for storage systems, or common abstractions, such as Ethernet service for
network devices. In addition, the logical driver for a single device may be split into two separate
drivers within the kernel, a high-level driver for interfacing with applications and a low-level driver
for interfacing with hardware. The benefit of the stacked model is that it simplifies extending the
capabilities of a device by inserting a filter driver between two layers in the stack. The difficulty
of stacked drivers, though, is that it encourages direct driver-to-driver communication. Drivers may
pass opaque pointers, including function pointers, amongst themselves. Without knowledge of the
types of these pointers, an isolation system cannot interpose on all communication or provide call-
by-value-result semantics. Hence, it cannot isolate the drivers from each other. In this model, either
the entire stack of drivers must be isolated in a single domain, or additional code is required for each
layer of the stack.
8.4 Memory Management
Much of the Nooks isolation code pertains to memory management. For example, the kernel invokes
Nooks whenever the kernel address space changes, in order to synchronize the page tables used by
lightweight kernel protection domains with the kernel page table. The object tracking code, to
provide call-by-value-result semantics, must be aware of how objects shared by drivers and the
kernel are allocated. As a result, how an operating system manages memory has a large impact on
the ease of isolating drivers.
Nooks provides isolation with lightweight kernel protection domains that map the kernel address
space with reduced access privileges. In a kernel that is not paged, the kernel address space changes
only when virtual mappings are needed to create large, contiguous memory blocks. Hence, when the
125
kernel page table changes, the shadow tables can be synchronized without affecting performance.
In a pageable kernel, such as Windows, changes to the address space are frequent, which leads
to higher costs to manage shadow page tables. In Linux, which is not paged, I chose to keep the
shadow and real page tables tightly synchronized. In a pageable kernel it may be more efficient to
update driver page tables on demand, because drivers access just a small subset of the entire kernel
address space.
The organization of the kernel memory management code itself also impacts the difficulty of
implementing lightweight kernel protection domains. To maintain tight synchronization of pages
tables with the kernel, Nooks must be notified whenever the kernel changes its page table. In
Linux, a small set of pre-processor macros manipulate the processor page tables. As a result, it is
straightforward to identify locations in the source code where the kernel updates page tables and
to insert synchronization code at this point. In contrast, Windows directly modifies hardware page
tables through union variables. As others have experienced [17], this design complicates the task of
identifying all code locations that manipulate page table contents. In our Windows implementation,
we chose to synchronize page tables when the TLB was flushed instead of tediously identifying
these locations.
The allocation patterns for data shared between the kernel and drivers also affect the ease of
isolation. The simplest pattern to isolate is when the kernel allocates a data object, passes it to the
driver, and the driver notifies the kernel when it is done accessing the object. In this pattern, the
life cycle of the object is clear: it has a clear birth (when it is allocated), use, and death (when it
is released). In contrast, it is more difficult to accurately track objects that are allocated on the ex-
tension’s stack. For example, Linux drivers pass stack-allocated objects to the kernel when waiting
for an event to occur. The object has no explicit birth, because it is created on the stack, and no
explicit death, because it dies when the stack is popped. Hence, it is difficult to track these objects
accurately. Nooks must instead track the lifetimes of these objects based on the kernel’s use of the
object; it creates the object when it is first passed to the kernel and deletes it when the kernel releases
it. This may be overly conservative, as the driver may reuse the same object multiple times. An anal-
ogous problem occurs in the Windows kernel. While Windows provides routines for allocating and
destroying I/O request packets (IRPs), drivers are also free to use their own allocation routines. It is
therefore difficult to detect whether an IRP is valid and when an IRP is dead, because a driver may
126
free it with many different routines. The object tracker conservatively frees the IRP whenever it is
unshared, which leads to unnecessary adding and removing of the same request packets from the
tracker.
The Nooks reliability system isolates drivers through memory protection. As a result, the com-
plexity of Nooks is strongly related to the complexity its hosting operating system’s memory man-
agement code. When the events of interest to Nooks, such as address space changes and object
creation and deletion, are explicit, then the isolation code is straightforward and efficient. When
these events are difficult to detect, then the isolation code is more complex and more costly. Fur-
thermore, the performance of Nooks depends on the frequency of such events. If address space
changes are frequent, as in paged kernels, then maintaining synchronized pages tables may become
prohibitively expensive and alternative isolation mechanisms, such as software fault isolation [205],
become more attractive.
8.5 Source Code Availability
The availability of driver source code proved important for two reasons. First, driver source code
provided details about the kernel-driver interface that are not documented, such as how drivers
manipulate kernel data structures. Second, the open-source development model encouraged Linux
to evolve towards a cleaner kernel-driver interface, which facilitates interposition, isolation and
recovery.
In implementing Nooks, I analyzed the source code from many device drivers. While device
driver code can potentially call any kernel function and manipulate any kernel data structure, this
analysis showed that, in practice, driver code is much more restrained. For example, of the 76 fields
in the net device structure for network device drivers, the drivers only write to a single field after
initialization. As a result of this analysis, the wrappers and object tracker avoid synchronizing the
other 75 fields on every entry and exit from a network driver. While this analysis may be possible
without extension source code, it would be far more difficult and error prone.
The source-code analysis was a snapshot of the behavior of a small number of drivers, and
may not remain valid as the kernel and drivers evolve. Once Nooks is incorporated into the kernel,
though, its existence pressures driver writers to follow Nooks assumptions: drivers that violate
127
the rules cannot be isolated or recovered automatically. Thus, in order to ensure that developers
conform to Nooks’ assumptions, it is crucial that Nooks achieve a critical mass early as a mandatory
component of the kernel and not an optional patch.
The open-source development model, in which driver code is available to be read and modified
by kernel developers, also had an impact on Nooks. From reading driver code across multiple kernel
versions, I observed a common pattern concerning the evolution of the kernel-driver interface. In
early versions of the kernel, many drivers directly modified kernel data structures, for example
changing flag fields in the current task structure. In later kernels, these accesses were codified into
pre-processor macros and inline functions. A simple change to the Linux header files therefore
allowed me to interpose on these functions and isolate the drivers with no changes to their source
code.
I believe that this process of abstracting common programming patterns is a positive feature
of open-source development. First, the driver code that accessed the kernel directly was visible
to kernel programmers and led them to extend the kernel interface. Thus, drivers use fewer so-
called “hacks” to access kernel resources, because kernel developers can formalize these hacks
and add them to the kernel-driver interface. Second, when modifying the kernel, developers could
change both the interface and the affected driver code. Thus, the open-source model allows a short
feedback loop between kernel and driver developers and leads to a cleaner interface, subject to
fewer work-arounds. In a closed-source environment, in contrast, the kernel developers may have
little knowledge of what driver writers are doing and have few means to encourage driver writers to
take advantage of new interfaces [155].
The availability of driver source code to kernel developers improves reliability by allowing ker-
nel developers to understand, at many levels, how drivers work and what drivers do. This leads to
cleaner kernel-driver interfaces that are subject to fewer work-arounds, which in turn lead to drivers
that are simpler to isolate and recover.
8.6 Summary
In summary, adding a reliability layer to an existing operating system requires injecting code on
all code paths that affect drivers. When these code paths are clearly identified, for example by
128
explicit function invocations, and easy to modify, then implementing a reliability subsystem can be
straightforward. If these events are widely distributed, difficulty to find, and difficult to detect, then
it is more difficult to tolerate driver failures.
Reliability also requires understanding the properties of the code that can fail, to tolerate the
failure and recover. When driver code is available, a reliability layer can make accurate assumptions
about the behavior of driver code. In addition, this code enables kernel developers to evolve the
kernel-driver interface in directions that simplify fault tolerance.
129
Chapter 9
FUTURE WORK
9.1 Introduction
While Nooks, shadow drivers, and dynamic driver update proved effective at improving system re-
liability and availability, there are many opportunities for future work. These opportunities fall into
two broad categories. First, the existing limitations of Nooks can be lifted or lessened and its per-
formance improved. For example, better analysis of device driver and extension code may remove
many of the compatibility problems mentioned in previous chapters. Second, new device driver
interfaces and architectures can lessen the substantial problems caused by device drivers. Drivers
must be rewritten for each operating system kernel and modified for each revision or update of de-
vice hardware. As a result, research operating system are often limited by the availability of drivers.
New driver architectures can lessen these problems by substantially reducing the dependency of
driver code on the hardware and the hosting OS.
9.2 Static Analysis of Drivers
The Nooks architecture makes strong assumptions about driver behavior. A driver that violates
these assumptions, while potentially bug-free, cannot run with Nooks. For example, wrappers allow
drivers to modify only a subset of the parameters they receive. A driver may fail at run-time if it
attempts to write to a read-only kernel object or if wrappers do not propagate changes it made to a
kernel object. In addition, shadow drivers require that device drivers implement the class interface
strictly and not store extra internal state based on device or kernel inputs. If a driver violates these
assumptions, then it may fail at run-time or may not recovery correctly.
These compatibility problems raise two opportunities to reduce the reliability impact of correct
code that Nooks cannot currently execute safely. First, tools and techniques that verify that a driver
is compatible with Nooks’ assumptions can ensure that all failures detected at run-time are real and
130
not caused by driver incompatibilities. Static analysis tools that operate on source code [68] or
binaries [12] could model check drivers against Nooks’ assumptions, such as which parameters are
read-only and which are read-write. By applying this check at install time, the system could avoid
isolating incompatible drivers.
The second opportunity for future work is to increase the range of drivers that are compatible
with Nooks. The existing Nooks implementation is based on a hand-analysis of a small number
of drivers. While this approach was sufficient for the drivers tested, automated analysis of a much
larger body of drivers would provide a more accurate model of driver behavior, and therefore allow
Nooks to support a wider variety of drivers. In addition, with automated analysis, it may be possi-
ble to create multiple models of drivers and find the best fit, i.e., one that minimizes the dynamic
overhead of copying data between protection domains. Again, static analysis can provide much of
the necessary information about how drivers use kernel objects. This analysis could also lead to
automated generation of wrapper, object-tracking.
A final use of static analysis techniques is to derive the interface and abstract state machine of a
driver automatically from its source code. Recent work on deriving interfaces from source code [3]
and on deriving interface properties from their implementations [69] could be applied to drivers.
Once found, this information can be used to model check new drivers and to generate shadow driver
code automatically.
9.3 Hybrid Static/Run-Time Checking
Nooks takes a purely fault-tolerance approach to reliability. Static analysis and safe languages
provide many, but not all, of the same benefits as Nooks. For example, the Vault language [58]
can verify statically that locking protocols are implemented correctly. Type-safe languages, such
as Java [88] and C] [99], ensure that code does not access invalid memory addresses. To date,
though, these techniques do not ensure that the driver provides the correct service, for example that
it correctly interfaces with hardware. Thus, even with static analysis of driver code, Nooks’ run-time
error detection and recovery services are still required for high availability.
Nonetheless, Nooks’ complexity and run-time overhead can be lessened by integrating these
fault avoidance techniques to reduce the amount of run-time checking. For example, drivers verified
131
statically do not require the memory isolation provided by lightweight kernel protection domains. In
addition, lock misuse need not be checked for drivers that have been verified to correctly implement
locking protocols. Thus, the run-time cost and complexity of Nooks can be reduced by checking
properties statically instead of at run-time.
9.4 Driver Design
Existing driver and extension architectures optimize for high performance, flexibility, and ease of
programming. For example, the Windows Driver Model supports three different memory manage-
ment models, allowing driver writers to balance performance and complexity [158]. The MacOS
I/O Kit provides an object-oriented framework for developing drivers, based on driver-class-specific
base classes, that reduces the amount of code needed for a specific driver [6].
However, these driver and extension architectures are not designed for reliability. First, they
ignore the cost of invoking an extension, a fundamental cost in any isolation or recovery system.
Second, they share kernel data at very fine granularity, complicating isolation. Third, they allow
opaque data to be passed between extensions, which prohibits marshaling or verifying of the data
contents. Finally, existing driver architectures require that all driver code run in the kernel, even
if it performs no privileged operations. For example, much of the code in network device drivers
concerns initialization and cleanup, error handling, and statistics gathering. None of these functions
are performance critical, and many of them require no privileged access to kernel data structures.
Hence, they need not execute with kernel privilege.
There several opportunities to rectify these limitations of the kernel-extension interface. First, all
invocations of extensions should be treated as if they were remote invocations, so that the parameters
are fully typed and could be marshaled if necessary. Second, the interface to the driver should
explicitly recognize the cost of invoking drivers and allow batching and asynchronous execution.
Third, driver code should be factored into code that must run in the kernel, for either performance or
privilege reasons, and code that need not. For example, driver initialization, cleanup, error handling,
and management code could run in a user-mode process instead of in the kernel. Thus, code in the
kernel that can potentially destabilize the system is kept to a minimum
Finally, there are opportunities for reducing the complexity of drivers that have not yet been fully
132
exploited by existing operating systems. While most operating systems implement abstractions for
devices that provide a common service, such as Ethernet drivers, or with a common command
protocol, such as SCSI disks, they do not take advantage of common hardware command interfaces.
For example, most network cards present a ring-buffer abstraction to their drivers. When sending
a packet, the driver places a pointer to the packet in the send ring and then signals the device that
it has updated the ring. Currently, every network driver re-implements the code to receive a packet
from the kernel, place it in the ring, and notify the device. The major difference between drivers is
the exact data format of the ring, such as the structure of the entries, and the commands for notifying
the device of new data. As a result, it may be possible to simplify drivers by providing abstractions
of hardware data structures, such as ring buffers, in the driver interface. Rather than writing the
entire send and receive path, the driver writer would instead provide a description of the ring data
structure and the command sequence. The OS can provide the rest of the code. The I/O system for
Mach 3.0 attempted a similar decomposition to abstract device-independent code from drivers [82].
The benefit of this approach is that it greatly simplifies the code for performing I/O, reducing the
cost of writing drivers and the risk of driver bugs.
An important aspect of evolving the driver interface is automatic migration from existing drivers
to the new framework. Recent work, such as the Ivy project for replacing C as a systems program
language [34], and Privtrans for automatically partitioning code for privilege separation [35], points
towards possible solutions to this problem.
There have been two recent projects that reinvent the driver interface. Project UDI (Uniform
Driver Interface) seeks to create a common driver interface across all Unix platforms [169]. The
design calls for fully asynchronous code with no blocking and no direct access to hardware. Instead,
blocking calls take callback functions and drivers access hardware though mini-programs written
in a domain-specific language. The Windows Driver Foundation seeks to simplify Windows device
drivers by providing a stripped-down calling interface and an interface definition language for driver
parameters [146]. Both efforts are a step in the right direction, but so far their adoption has been
hampered by lack of compatibility with existing driver code.
133
9.5 Driverless Operating Systems
Beyond redesigning drivers, it is possible to imagine operating systems that do away with device
drivers altogether. Even with an improved interface, drivers would continue to pose problems: they
must be updated for every device revision, they must be ported to every operating system, and
they are critical to performance and reliability. As a result, drivers will continue to hamper system
development unless a new solution is found.
Drivers today serve as a combination of a translation layer, connecting the operating system
interface to the device interface, an adaptation layer, masking the impedance mismatch between
processor and device speeds, and a service layer, providing functionality beyond the device’s abili-
ties. However, these three roles may be filled by different software that need not suffer from driver
problems. Existing safe-extension technology can provide the service layer functionality, but new
communication abstractions are needed for the other two functions.
The job of interfacing to device hardware is necessary because higher-level communication
abstractions do not exist for drivers. Elsewhere in operating system design, these abstractions have
simplified application design. For example, remote procedure call systems remove much of the
complexity in implementing client-server communication. In contrast, many drivers still access
devices by reading and writing memory-mapped device registers. While new device connection
technologies, such as FireWire [62] and USB [49], allow drivers and devices to communicate with
messages, the richness of distributed programming techniques is not yet available to drivers.
To address this issue, new communications mechanisms that take into account the needs of de-
vices, such as low latency, high bandwidth, and direct access to memory, are needed. These mecha-
nisms must incorporate the costs of communication and the relative speeds of the device and the host
processor to serve the impedance-matching role of today’s drivers. While most existing approaches
to programming distributed systems do not apply to device drivers directly, the abstractions, lessons,
and approaches can greatly simplify driver design. For example, Devil’s use of interface definition
languages replaces much of the handcrafted code in drivers that communicates with devices [142].
High-performance messaging abstractions, such as Active Messages [203] may also apply.
An important benefit of high-level communication protocols is that it pushes driver function-
ality out of the operating system, where it impacts system reliability and must be ported or re-
134
implemented for each additional OS, onto the hardware itself. Thus, much of the code in drivers
need be implemented just once, on the driver hardware, rather than separately for each operating
system.
With these techniques, operating systems will not need drivers as we conceive of them today. In-
stead, the function of drivers can be shifted to user-level services, application libraries, and hardware
devices. Systems will invoke driver services over a common communication substrate, simplifying
the OS and improving its reliability. The code that currently constitutes drivers will be easier to
update as new devices are released, because it relies on high-level mechanisms instead of low-level
device access. Finally, this code can be ported between operating systems more easily because it
can be executed outside of the kernel, as an application library or user-level service.
9.6 Summary
There are two major directions to pursue future research. First, the existing architecture of Nooks
can be extended to cover a greater fraction of kernel extensions and to incorporate static checking as
a substitute for run-time checking. With these technologies, Nooks isolation and recovery services
could be applied to other extensible systems that depend on extension correctness for reliability.
Second, the lessons from implementing Nooks point toward new considerations for device driver
design. A new interface to drivers, tailored to reliability concerns, could greatly reduce driver-
induced failures. Longer term, new hardware designs and new communication mechanisms could
lead to operating systems without device drivers as we know them.
135
Chapter 10
CONCLUSIONS
Reliability has become a critical challenge for commodity operating systems as we depend on
computers for more of our daily lives. The competitive pressure on these systems and their huge
installed base, though, prevents the adoption of traditional fault-tolerance techniques.
This thesis presents a new approach to improving the reliability of operating systems that is at
once efficient and backwards compatible. Rather than tolerate all possible failures, my approach
is to target the most common failures and thereby improve reliability at very low cost. In today’s
commodity operating systems, device driver failures are the dominant cause of system failure.
In this thesis, I described Nooks, a new reliability layer intended to prevent drivers from forcing
either the OS or applications to restart. Nooks uses hardware and software techniques to isolate
device drivers, trapping many common faults and permitting extension recovery. Shadow drivers
ensure that the OS and applications continue to function during and after recovery. Dynamic driver
update ensures that applications and the OS continue to run when applying driver updates.
The Nooks system accomplishes these tasks by focusing on backward compatibility. That is,
Nooks sacrifices complete isolation and fault tolerance for compatibility and transparency with ex-
isting kernels and drivers. Nevertheless, Nooks demonstrates that it is possible to realize an ex-
tremely high level of operating system reliability with low performance lost for common device
drivers. My fault-injection experiments show that Nooks recovered from 99% of the faults that
caused native Linux to crash. Shadow drivers kept applications running for 98% of the faults that
would otherwise have caused a failure. In testing, dynamic driver update could apply 99% of driver
updates without rebooting the system or restarting applications. Furthermore, the performance cost
of Nooks averaged only 1% across common applications and drivers.
My experience shows that: (1) implementation of a Nooks layer is achievable with only modest
engineering effort, even on a monolithic operating system like Linux, (2) device drivers can be
isolated without change to driver code, and (3) isolation and recovery can dramatically improve the
136
system’s ability to survive extension faults.
Looking forwards, Nooks demonstrates that reliability should be a first-class citizen when de-
signing extension interfaces. In addition to performance and programmability, interfaces should be
designed to enable efficient isolation and recovery by exposing key events and establishing clear
boundaries between extensions and the core system. Only when this lesson has been learned will
extensible systems, such as operating systems, browsers, and web servers, achieve high reliability.
137
BIBLIOGRAPHY
[1] Edward N. Adamas. Optimizing preventive maintenance of software products. IBM Journalof Research and Development, 28(1):2–14, January 1984.
[2] Gautam Altekar, Ilya Bagrak, Paul Burstein, and Andrew Schultz. OPUS: Online patchesand updates for security. In Proceedings of the 14th USENIX Security Symposium, pages287–302, Baltimore, Maryland, August 2005.
[3] Rajeev Alur, Pavol Cerny, P. Madhusudan, and Wonhong Nam. Synthesis of interface spec-ifications for java classes. In Proceedings of the 32nd ACM SIGPLAN-SIGACT Symposiumon Principles of Programming Languages, pages 98–109, Long Beach, California, January2005.
[4] Scott Maxwell an Frank Hartman. Linux and JPL’s mars exploratin rover project: Earth-basedplanning, simulation, and really remote scheduling. USENIX Annual Technical ConferenceInvited Talk, April 2005.
[5] Apache Project. Apache HTTP server version 2.0, 2000. Available at http://httpd.apache.org.
[6] Apple Computer. I/O kit fundamentals, 2005. Available at http://developer.apple.com/documentation/DeviceDrivers/Conceptual/IOKitFundamentals/IOKitFundamentals.pdf.
[7] Hitoshi Araki, Shin Futagami, and Kaoru Nitoh. A non-stop updating technique for devicedriver programs on the IROS platform. In IEEE International Conference on Communica-tions, pages 18–22, Seattle, WA, June 1995.
[8] Remzi H. Arpaci-Dusseau and Andrea C. Arpaci-Dusseau. Fail-stutter fault tolerance. InProceedings of the Eighth IEEE Workshop on Hot Topics in Operating Systems, pages 33–40, Schloss Elmau, Germany, May 2001.
[9] Sandy Arthur. Fault resilient drivers for Longhorn server. Microsoft WinHec 2004 Presenta-tion DW04012, May 2004.
[10] Algirdas Avizienis, Jean-Claude Laprie, Brian Randell, and Carl Landwehr. Basic conceptsof dependable and secure computing. IEEE Transactions on Dependable and Secure Com-puting, 1(1):11–33, January 2004.
138
[11] Ozalp Babaoglu. Fault-tolerant computing based on Mach. In Proceedings of the USENIXMach Symposium, pages 185–199, Burlington, VT, October 1990.
[12] Gogul Balakrishnan, Radu Gruian, Thomas R. Reps, and Tim Teitelbaum. CodeSurfer/x86 –a platform for analyzing x86 executables. In Proceedings of the International Conference onCompiler Construction, pages 250–254, Edinburgh, Scotland, April 2005.
[13] Thomas Ball and Sriram K. Rajamani. Automatically validating temporal safety properties ofinterfaces. In SPIN 2001, Workshop on Model Checking of Software, volume 2057 of LectureNotes in Computer Science, pages 103–122, May 2001.
[14] Thomas Ball and Sriram K. Rajamani. The SLAM project: Debugging system software viastatic analysis. In Proceedings of the 29th ACM SIGPLAN-SIGACT Symposium on Principlesof Programming Languages, pages 1–3, Portland, OR, January 2002.
[15] Roger Barga, David Lomet, German Shegalov, and Gerhard Weikum. Recovery guaranteesfor internet applications. ACM Transactions on Internet Technology, 2004.
[16] Roger Barga, David Lomet, and Gerhard Weikum. Recovery guarantees for general multi-tierapplications. In International Conference on Data Engineering, pages 543–554, San Jose,California, 2002. IEEE.
[17] Paul Barham, Boris Dragovic, Keir Fraser, Steven Hand, Tim Harris, Alex Ho, Rolf Neuge-bauer, Ian Pratt, and Andrew Warfield. Xen and the art of virtualization. In Proceedings ofthe 19th ACM Symposium on Operating Systems Principles, pages 164–177, Bolton Landing,New York, October 2003.
[18] Mike Barnett, K. Rustan, M. Leino, and Wolfram Schulte. The Spec] programming system:An overview. In CASSIS International Workshop, volume 3362 of Lecture Notes in ComputerScience, pages 49–69, Marseille, France, March 2004. Springer-Verlag.
[19] Joel F. Bartlett. A NonStop kernel. In Proceedings of the 8th ACM Symposium on OperatingSystems Principles, pages 22–29, Pacific Grove, California, December 1981.
[20] Elisa Batista. Windows to power ATMs in 2005. Wired News, September 2003. http://www.wired.com/news/technology/0,1282,60497,00.html.
[21] Andrew Baumann and Gernot Heiser. Providing dynamic update in an operating system. InProceedings of the 2005 USENIX Annual Technical Conference, pages 297–291, Anaheim,California, April 2005.
[22] David Bernick, Bill Bruckert, Paul Del Vigna, David Garcia, Robert Jardine, Jim Klecka, andJim Smullen. NonStop advanced architecture. In Proceedings of the 2005 IEEE InternationalConference on Dependable Systems and Networks, pages 12–21, Yokohama, Japan, June2005.
139
[23] Brian N. Bershad, Thomas E. Anderson, Edward D. Lazowska, and Henry M. Levy.Lightweight remote procedure call. ACM Transactions on Computer Systems, 8(1):37–55,February 1990.
[24] Brian N. Bershad, Stefan Savage, Przemysław Pardyak, Emin Gun Sirer, Marc E. Fiuczynski,David Becker, Craig Chambers, and Susan Eggers. Extensibility, safety and performancein the SPIN operating system. In Proceedings of the 15th ACM Symposium on OperatingSystems Principles, pages 267–284, Copper Mountain, Colorado, December 1995.
[25] Andrew D. Birrell and Bruce J. Nelson. Implementing remote procedure calls. ACM Trans-actions on Computer Systems, 2(1):39–59, February 1984.
[26] Common Criteria Implementation Board. Common criteria for information technology se-curity evaluation, part 3: Security assurance requirements, version 2.2. Technical ReportCCIMB-2004-01-003, National Institute of Standards and Technology, January 2004.
[27] Aniruddha Bohra, Iulian Neamtiu, Pascal Gallard, Florin Sultan, and Liviu Iftode. Remoterepair of operating system state using Backdoors. In Proceedings of the International Con-ference on Autonomic Computing, pages 256–263, New York, New York, May 2004.
[28] Anita Borg, Wolfgang Balu, Wolfgang Graetsch, Ferdinand Herrmann, and Wolfgang Oberle.Fault tolerance under UNIX. ACM Transactions on Computer Systems, 7(1):1–24, February1989.
[29] David A. Borman. Implementing TCP/IP on a Cray computer. ACM SIGCOMM ComputerCommunication Review, 19(2):11–15, April 1989.
[30] Luciano Porto Borreto and Gilles Muler. Bossa: A language-based approach for the designof real time schedulers. In Proceedings of the 14th Euromicro Conference on Real TimeSystems, Paris, France, March 2002.
[31] Daniel P. Bovet and Marcho Cesati. Understanding the Linux Kernel, 2nd Edition. O’ReillyAssociates, December 2002.
[32] Thomas C. Bressoud. TFT: A software system for application-transparent fault tolerance. InProceedings of the 28th Symposium on Fault-Tolerant Computing, pages 128–137, Munich,Germany, June 1998. IEEE.
[33] Thomas C. Bressoud and Fred B. Schneider. Hypervisor-based fault tolerance. ACM Trans-actions on Computer Systems, 14(1):80–107, February 1996.
[34] Eric Brewer, Jeremy Condit, Bill McCloskey, and Feng Zou. Thirty years is long enough:Getting beyond C. In Proceedings of the Tenth IEEE Workshop on Hot Topics in OperatingSystems, Santa Fe, New Mexico, June 2005.
140
[35] David Brumley and Dawn Song. Privtrans: Automatically partitioning programs for privi-lege separation. In Proceedings of the 13th USENIX Security Symposium, pages 57–72, SanDiego, California, August 2004.
[36] George Candea and Armando Fox. Recursive restartability: Turning the reboot sledgeham-mer into a scalpel. In Proceedings of the Eighth IEEE Workshop on Hot Topics in OperatingSystems, pages 125–132, Schloss Elmau, Germany, May 2001.
[37] Miguel Castro and Barbara Liskov. Practical byzantine fault tolerance. In Proceedings of the3rd USENIX Symposium on Operating Systems Design and Implementation, pages 173–186,New Orleans, Louisiana, February 1999.
[38] Subhachandra Chandra and Peter M. Chen. How fail-stop are faulty programs? In Proceed-ings of the 28th Symposium on Fault-Tolerant Computing, pages 240–249, Munich, Germany,June 1998. IEEE.
[39] John Chapin, Mendel Rosenblum, Scott Devine, Tirthankar Lahiri, Dan Teodosiu, and AnoopGupta. Hive: Fault containment for shared-memory multiprocessors. In Proceedings ofthe 15th ACM Symposium on Operating Systems Principles, pages 12–25, Copper MountainResort, Colorado, December 1995.
[40] Jeffrey S. Chase, Henry M. Levy, Michael J. Feeley, and Edward D. Lazowska. Sharingand protection in a single-address-space operating system. ACM Transactions on ComputerSystems, 12(4):271–307, November 1994.
[41] L. Chen and Algirdas Avizienis. N-version programming: A fault-tolerance approach toreliability of software operation. In Proceedings of the 8th Symposium on Fault-TolerantComputing, pages 3–9, Toulouse, France, June 1978.
[42] Mike Y. Chen, Anthony Accardi, Emre Kiciman, Dave Patterson Jim Lloyd, Armando Fox,and Eric Brewer. Path-based failure and evolution management. In Proceedings of the 2004USENIX Symposium on Network Systems Design and Implementation, pages 309–322, SanFrancisco, California, March 2004.
[43] Peter M. Chen and Brian Noble. When virtual is better than real. In Proceedings of theEighth IEEE Workshop on Hot Topics in Operating Systems, pages 133–138, Schloss Elmau,Germany, May 2001.
[44] Tzi-cker Chiueh, Ganesh Venkitachalam, and Prashant Pradhan. Integrating segmentationand paging protection for safe, efficient and transparent software extensions. In Proceedingsof the 17th ACM Symposium on Operating Systems Principles, pages 140–153, Kiawah IslandResort, South Carolina, December 1999.
141
[45] Andy Chou, Junfeng Yang, Benjamin Chelf, Seth Hallem, and Dawson Engler. An empiricalstudy of operating system errors. In Proceedings of the 18th ACM Symposium on OperatingSystems Principles, pages 73–88, Lake Louise, Alberta, October 2001.
[46] Jorgen Christmansson and Ram Chillarege. Generation of an error set that emulates soft-ware faults - based on field data. In Proceedings of the 26th Symposium on Fault-TolerantComputing, pages 304 – 313, Sendai, Japan, June 1996. IEEE.
[47] Yvonne Coady and Grigor Kiczales. Back to the future: A retroactive study of aspect evo-lution in operating systems code. In Proceedings of the International Conference on Aspect-Oriented Software Development, pages 50–59, Boston, Massachusetts, March 2003.
[48] Robert P. Colwell, Edward F. Gehringer, and E. Douglas Jensen. Performance effects ofarchitectural complexity in the Intel 432. ACM Transactions on Computer Systems, 6(3):296–339, August 1988.
[49] Compaq, Hewlett-Packard, Intel, Lucent, Microsoft, NEC, and Philips. Universal serial bus(USB) 2.0 specification, April 2000. Available at http://www.usb.org/developer/docs.
[50] Jeremy Condit, Matthew Harren, Scott McPeak, George C. Necula, and Westley Weimer.CCured in the real world. In Proceedings of the ACM SIGPLAN ’03 Conference on Program-ming Language Design and Implementation, pages 232–244, San Diego, California, USA,June 2003.
[51] Fernando J. Corbato and V. A. Vyssotsky. Introduction and overview of the multics system.In Proceedings of the AFIPS Fall Joint Computer Conference, volume 27, pages 185–196,Las Vegas, Nevada, 1965. Spartan Books.
[52] Jonathan Corbet, Alessandro Rubini, and Greg Kroah-Hartman. Linux Device Drivers, 3rdEdition. O’Reilly Associates, February 2005.
[53] Coverity. Anaylsis of the Linux kernel, 2004. Available at http://www.coverity.com.
[54] Flaviu Cristian. Understanding fault-tolerant distributed systems. Communications of theACM, 34(2):56–78, February 1991.
[55] Helen Custer. Inside Windows NT. Microsoft Press, Redmond, WA, 1993.
[56] D. H. Brown Associates. Stratus Introduces Four-Way, Fault-Tolerant SMP Server withHyper-Threading for Windows 2000 Advanced Server, September 2002. Available athttp://www.stratus.com/resources/pdf/brown092002.pdf.
[57] Joseph Dadzie. Understanding software patching. ACM Queue, 3(2), March 2005.
142
[58] Robert DeLine and Manuel Fahndrich. Enforcing high-level protocols in low-level software.In Proceedings of the ACM SIGPLAN ’01 Conference on Programming Language Designand Implementation, pages 59–69, Snowbird, Utah, June 2001.
[59] Peter J. Denning. Fault tolerant operating systems. ACM Computing Surveys, 8(4):359–389,December 1976.
[60] Jack B. Dennis and Earl Van Horn. Programming semantics for multiprogramming systems.Communications of the ACM, 9(3), March 1966.
[61] David J. DeWitt. The Wisconsin benchmark: Past, present, and future. In Jim Gray, editor,The Benchmark Handbook for Database and Transaction Systems, Second Edition, pages269–315. Morgan Kaufmann, 1993.
[62] Mindshare Inc. Don Anderson. FireWire System Architecture: IEEE 1394A. Addison-WesleyProfessional, 1999.
[63] R. H. Doyle, R. A. Meyer, and R. P. Pedowitz. Automatic failure recovery in a digital dataprocessing system. IBM Journal of Research and Development, 3(1):2–12, January 1959.
[64] Richard P. Draves. The case for run-time replaceable kernel modules. In Proceedings ofthe Proceedings of the Fourth Workshop on Workstation Operating Systems, pages 160–164,Napa, California, October 1993. IEEE Computer Society Press.
[65] Richard P. Draves, Brian N. Bershad, Richard F. Rashid, and Randall W. Dean. Using con-tinuations to implement thread management and communications in operating systems. InProceedings of the 13th ACM Symposium on Operating Systems Principles, pages 122–136,Pacific Grove, California, October 1991.
[66] Dominic Duggan. Type-based hot swapping of running modules. Technical Report SIT CSReport 2001-7, Stevens Institute of Technology, Castle Point on the Hudson, Hoboken, NewJersey, October 2001.
[67] Gregory Duval and Jacques Julliand. Modeling and verification of the RUBIS µ-kernel withSPIN. In Proceedings of the International SPIN Workshop, Montreal, Canada, October 1995.
[68] Dawson Engler, Benjamin Chelf, Andy Chou, and Seth Hallem. Checking system rulesusing system-specific, programmer-written compiler extensions. In Proceedings of the 4thUSENIX Symposium on Operating Systems Design and Implementation, pages 1–16, SanDiego, California, October 2000.
[69] Dawson Engler, David Yu Chen, and Andy Chou. Bugs as inconsistent behavior: A generalapproach to inferring errors in systems code. In Proceedings of the 18th ACM Symposium onOperating Systems Principles, pages 57–72, Lake Louise, Alberta, October 2001.
143
[70] Dawson R. Engler. personal communication, March 2005.
[71] Dawson R. Engler, M. Frans Kaashoek, and James O’Toole Jr. Exokernel: an operatingsystem architecture for application-level resource management. In Proceedings of the 15thACM Symposium on Operating Systems Principles, pages 251–266, Copper Mountain Resort,Colorado, December 1995.
[72] Ulfar Erlingsson, Tom Roeder, and Ted Wobber. Virtual environments for unreliable exten-sions. Technical Report MSR-TR-05-82, Microsoft Research, June 2005.
[73] Jean-Charles Fabre, Manuel Rodrı, Jean Arlat, Frederic Salles, and Jean-Michel Sizun. Build-ing dependable COTS microkernel-based systems using MAFALDA. In Proceedings of the2000 Pacific Rim International Symposium on Dependable Computing, pages 85–94, LosAngeles, California, December 2000.
[74] Robert S. Fabry. Capability-based addressing. Communications of the ACM, 17(7):403–412,July 1974.
[75] Robert S. Fabry. How to design a system in which modules can be changed on the fly. InProceedings of the 2nd International Conference on Software Engineering, pages 470–476,October 1976.
[76] Michael E. Fagan. Design and code inspections to reduce errors in program development.IBM Systems Journal, 15(3):182–211, 1976.
[77] Wu-Chun Feng. Making a case for efficient supercomputing. ACM Queue, 1(7), October2003.
[78] Christof Fetzer and Zhen Xiao. HEALERS: A toolkit for enhancing the robustness and secu-rity of existing applications. In Proceedings of the 2003 IEEE International Conference onDependable Systems and Networks, pages 317–322, San Francisco, California, June 2003.IEEE.
[79] Charles Fishman. They write the right stuff. Fast Company, (6):95–103, December 1996.
[80] Marc E. Fiuczynski, Robert Grimm, Yvonne Coady, and David Walker. patch(1) consideredharmful. In Proceedings of the Tenth IEEE Workshop on Hot Topics in Operating Systems,Santa Fe, New Mexico, June 2005.
[81] Bryan Ford, Godmar Back, Greg Benson, Jay Lepreau, Albert Lin, and Olin Shivers. TheFlux OSKit: a substrate for OS language and research. In Proceedings of the 16th ACMSymposium on Operating Systems Principles, pages 38–51, October 1997.
144
[82] Alessandro Forin, David Golub, and Brian Bershad. An I/O system for Mach. In Proceedingsof the Usenix Mach Symposium, pages 163–176, November 1991.
[83] Keir Fraser, Steven Hand, Rolf Neugebauer, Ian Pratt, Andrew Warfield, and MarkWilliamson. Safe hardware access with the Xen virtual machine monitor. In Proceedingsof the Workshop on Operating System and Architectural Support for the On- Demand ITInfrastructure, October 2004.
[84] Michael Gagliardi, Ragunathan Rajkumar, and Lui Sha. Designing for evolvability: Buildingblocks for evolvable real-time systems. In Proceedings of the 2nd IEEE Real-Time Technol-ogy and Applications Symposium, pages 100–109, Boston, Massachusetts, June 1996.
[85] James Gettys, Philip L. Carlton, and Scott McGregor. The X window system version 11.Technical Report CRL-90-08, Digital Equipment Corporation, December 1900.
[86] Al Gillen, Dan Kusnetzky, and Scott McLaron. The role of Linux in reducing the cost ofenterprise computing, January 2002. IDC white paper.
[87] Michael Golm, Meik Felser, Christian Wawersich, and Jurgen Kleinoder. The JX operatingsystem. In Proceedings of the 2002 USENIX Annual Technical Conference, pages 45–58,Monterey, California, June 2002.
[88] James Gosling, Bill Joy, and Guy Steele. The Java Language Specification. Addison-Wesley,1996.
[89] Hannes Goullon, Rainer Isle, and Klaus-Peter Lohr. Dynamic restructuring in an experi-mental operating system. In Proceedings of the 3rd International Conference on SoftwareEngineering, pages 295–304, May 1978.
[90] Jim Gray. The transaction concept: Virtues and limitations. In Proceedings of the 7th In-ternational Conference on Very Large Data Bases, pages 144–154, Los Angeles, California,September 1981.
[91] Jim Gray. Why do computers stop and what can be done about it? In Proceedings of theFifth Symposium on Reliability in Distributed Software and Database Systems, pages 3–12,Los Angeles, California, January 1996. IEEE.
[92] Jim Gray. Heisenbugs: A probabilistic approach to availability. Talk given at ISAT,availble at http://research.microsoft.com/∼Gray/talks/ISAT Gray FTAvialiability talk.ppt, April 1999.
[93] Jim Gray and Daniel P. Siewiorek. High-availability computer systems. IEEE Computer,24(9):39–48, September 1991.
145
[94] Deepak Gupta and Pankaj Jalote. On-line software version change using state transfer be-tween processes. Software–Practice and Experience, 23(9):949–94, September 1993.
[95] Jaap C. Haarsten. The Bluetooth radio system. IEEE Personal Communications Magazine,7(1):28–36, February 2000.
[96] Steven M. Hand. Self-paging in the Nemesis operating system. In Proceedings of the 3rdUSENIX Symposium on Operating Systems Design and Implementation, pages 73–86, NewOrleans, Lousiana, February 1999.
[97] Hermann Hartig, Michael Hohmuth, Jochen Liedtke, Sebastian Schoberg, and Jean Wolter.The performance of µ-kernel-based systems. In Proceedings of the 16th ACM Symposium onOperating Systems Principles, pages 66–77, Saint-Malo, France, October 1997.
[98] Steffen Hauptmann and Josef Wasel. On-line maintenance with on-the-fly software replace-ment. In Proceedings of the 3rd International Conference on Configurable Distributed Sys-tems, pages 70–80, May 1996.
[99] Anders Hejlsberg, Scott Wiltamuth, and Peter Golde. The C] Programming Language.Addison-Wesley, Boston, Massachusetts, October 2003.
[100] Thomas A. Henzinger, Ranjit Jhala, Rupak Majumdar, George C. Necula, Gregoire Sutre,and Westley Weimer. Temporal-safety proofs for systems code. In Proceedings of the 14thInternational Conference on Computer-Aided Verification (CAV), volume 2404 of LectureNotes in Computer Science, pages 526–538. Springer-Verlag, July 2002.
[101] Scott Herboldt. How to improve driver quality with Winqual/WHQL. Microsoft WinHec2005 Presentation TWDE05006, April 2005.
[102] Hewlett Packard. Hewlett Packard Digital Entertainment Center, October 2001. http://www.hp.com/hpinfo/newsroom/press/31oct01a.htm.
[103] Michael Hicks, Jonathan T. Moore, and Scott Nettles. Dynamic software updating. In Pro-ceedings of the ACM SIGPLAN ’01 Conference on Programming Language Design and Im-plementation, pages 13–23, Snowbird, Utah, June 2001.
[104] Dan Hildebrand. An architectural overview of QNX. In Proceedings of the USENIX Work-shop on Microkernels and Other Kernel Architectures, pages 113–126, Seattle, Washington,1992.
[105] Gısli Hjalmtysson and Robert Gray. Dynamic C++ classes: A lightweight mechanism toupdate code in a running program. In Proceedings of the 1998 USENIX Annual TechnicalConference, pages 65–76, New Orleans, Louisiana, June 1998.
146
[106] Gerald J. Holzmann. The model checker SPIN. IEEE Transactions on Software Engineering,23(5), May 1997.
[107] Merle E. Houdek, Frank G. Soltis, and Roy L. Hoffman. IBM System/38 support forcapability-based addressing. In Proceedings of the 8th ACM International Symposium onComputer Architecture, pages 341–348, Minneapolis, Minnesota, May 1981.
[108] Mei-Chen Hsueh, Timothy K. Tsai, and Ravishankar K. Iyer. Fault injection techniques andtools. IEEE Computer, 30(4):75–82, April 1997.
[109] Galen Hunt. Creating user-mode device drivers with a proxy. In Proceedings of the 1997USENIX Windows NT Workshop, Seattle, WA, August 1997.
[110] Galen C. Hunt and James R. Larus. Singularity design motivation (Singularity TechnicalReport 1). Technical Report MSR-TR-2004-105, Microsoft Research, December 2004.
[111] RTCA Inc. Software considerations in airborne systems and equipment certification, Decem-ber 1992. RTCA/DO-178B.
[112] Intel Corporation. The IA-32 Architecture Software Developer’s Manual, Volume 1: BasicArchitecture. Intel Corporation, January 2002. Available at http://www.intel.com/design/pentium4/manuals/24547010.pdf.
[113] Doug Jewett. Integrity S2: A fault-tolerant Unix platform. In Proceedings of the 21st Sym-posium on Fault-Tolerant Computing, pages 512–519, Quebec, Canada, June 1991. IEEE.
[114] Trevor Jim, Greg Morrisett, Dan Grossman, Michael Hicks, James Cheney, and YanlingWang. Cyclone: A safe dialect of C. In Proceedings of the 2002 USENIX Annual Tech-nical Conference, pages 275–288, Monterey, California, June 2002.
[115] Andreeas Johansson and Neeraj Suri. Error propagation profiling of operating systems. InProceedings of the 2005 IEEE International Conference on Dependable Systems and Net-works, pages 86–95, Yokohama, Japan, June 2005.
[116] Rick Jones. Netperf: A network performance benchmark, version 2.1, 1995. Available athttp://www.netperf.org.
[117] Jungo. Windriver cross platform device driver development environment. Technical report,Jungo Corporation, February 2002. http://www.jungo.com/windriver.html.
[118] Gregor Kiczales, John Lamping, Anurag Mendhekar, Chris Maeda, Cristina Videria Lopes,Jean-Marc Loingtier, and John Irwin. Aspect-oriented programming. In Proceedings of the11th European Conference on Object-Oriented Programming, volume 1241 of Lecture Notesin Computer Science, pages 220–242, Jyvasklya, Finland, June 1997. Springer-Verlag.
147
[119] Mark J. Kilgard, David Blythe, and Deanna Hohn. System support for OpenGL direct ren-dering. In Proceedings of Graphics Interface, pages 116–127, Toronto, Ontario, May 1995.Canadian Human-Computer Communications Society.
[120] Arun Kishan and Monica Lam. Dynamic kernel modification and extensibility, May 2002.http://suif.stanford.edu/papers/kishan.pdf.
[121] John C. Knight and Nancy G. Leveson. An experimental evaluation of the assumption ofindependence in multiversion programming. IEEE Transactions on Software Engineering,12(1):96–109, January 1986.
[122] Michael Kofler. The Definitive Guide to mySQL, Second Edition. Apress, Berkeley, Califor-nia, October 2003.
[123] Eric J. Koldinger, Jefrey S. Chase, and Susan J. Eggers. Architectural support for singleaddress space operating systems. In Proceedings of the Fifth ACM International Conferenceon Architectural Support for Programming Languages and Operating Systems, pages 175–186, Boston, Massachusetts, October 1992.
[124] Sanjeev Kumar and Kai Li. Using model checking to debug device firmware. In Proceedingsof the 5th USENIX Symposium on Operating Systems Design and Implementation, pages 61–74, Boston, Massachusetts, December 2002.
[125] Jean-Claude Laprie, Jean Arlat, Christian Beounes, Kerama Kanoun, and Catherine Hour-tolle. Hardware and software fault tolerance: Definition and analysis of architectural solu-tions. In Proceedings of the 17th Symposium on Fault-Tolerant Computing, pages 116–121,Pittsburgh, Pennsylvania, July 1987. IEEE.
[126] James R. Larus, Thomas Ball, Manuvir Das, Robert DeLine, Manuel Fahndrich, Jon Pincusadn Sriram K. Rajamani, and Ramanathan Venkatapathy. Righting software. IEEE Transac-tions on Software Engineering, 30(5):92–100, May 2004.
[127] Julia L. Lawall, Gilles Muller, and Richard Urunuela. Tarantula: Killing driver bugs beforethey hatch. In Proceedings of the Fourth AOSD Workhop on Aspects, Components, andPatterns for Infrastructure Software, pages 13–18, Chicago, Illinois, March 2005.
[128] Inhwan Lee and Ravishankar K. Iyer. Software dependability in the Tandem GUARDIANsystem. IEEE Transactions on Software Engineering, 21(5), May 1995.
[129] Insup Lee. DYMOS: A Dynamic Modification System. PhD thesis, Department of ComputerScience, University of Wisconsin, Madison, Wisconsin, April 1983.
148
[130] Joshua LeVasseur, Volkmar Uhlig, Jan Stoess, and Stefan Go. Unmodified device driver reuseand improved system dependability via virtual machines. In Proceedings of the 6th USENIXSymposium on Operating Systems Design and Implementation, pages 17–30, San Francisco,California, December 2004.
[131] Henry M. Levy. Capability-Based Computer Systems. Digital Press, 1984. Available athttp://www.cs.washington.edu/homes/levy/capabook.
[132] Jochen Liedtke. On µ-kernel construction. In Proceedings of the 15th ACM Symposium onOperating Systems Principles, pages 237–250, Copper Mountain Resort, Colorado, Decem-ber 1995.
[133] Linux Kernel Mailing List. Available at http://www.uwsg.indiana.edu/hypermail/linux/kernel.
[134] Bev Littlewood and Lorenzo Strigini. Software reliablity and dependability: a roadmap. InAnthony Finkelstein, editor, Proceedings of the International Conference on Software En-gineering - Future of Software Engineering Track, pages 175–188, Limerick, Ireland, June2000.
[135] David E. Lowell, Subhachandra Chandra, and Peter M. Chen. Exploring failure transparencyand the limits of generic recovery. In Proceedings of the 4th USENIX Symposium on Oper-ating Systems Design and Implementation, pages 289–304, San Diego, California, October2000.
[136] David E. Lowell and Peter M. Chen. Discount checking: Transparent, low-overhead re-covery for general applications. Technical Report CSE-TR-410-99, University of Michigan,November 1998.
[137] David E. Lowell, Yasushi Saito, and Eileen J. Samberg. Devirtualizable virtual machinesenabling general, single-node, online maintenance. In Proceedings of the Eleventh ACM In-ternational Conference on Architectural Support for Programming Languages and OperatingSystems, pages 211–223, October 2004.
[138] LWN.net. Kernel patch page. Available at http://lwn.net/Kernel/Patches.
[139] R. E. Lyons and W. Vanderkulk. The use of triple modular redundancy to improve computerreliability. IBM Journal of Research and Development, 7(2):200–209, April 1962.
[140] Sean McGrane. Windows Server platform: Overview and roadmap. Microsoft WinHec 2005Presentation TWSE05002, April 2005.
149
[141] Paul R. McJones and Garret F. Swart. Evolving the unix system interface to support multi-threaded programs. Technical Report SRC Research Report 21, Digital Systems ResearchCenter, September 1987. Available at http://gatekeeper.research.compaq.com/pub/DEC/SRC/research-reports/SRC-021-html/evolve.html.
[142] Fabrice Merillon, Laurent Reveillere, Charles Consel, Renaud Marlet, and Gilles Muller.Devil: An IDL for hardware programming. In Proceedings of the 4th USENIX Symposium onOperating Systems Design and Implementation, pages 17–30, San Diego, California, October2000.
[143] Microsoft Corporation. Windows Update. Available at http://www.windowsupdate.com.
[144] Microsoft Corporation. FAT: General overview of on-disk format, version 1.03, December2000.
[145] Microsoft Corporation. Cancel logic in windows drivers, May 2003. Available at http://www.microsoft.com/whdc/driver/kernel/cancel logic.mspx.
[146] Microsoft Corporation. Architecture of the windows driver foundation, July 2005. Availableat http://www.microsoft.com/whdc/driver/wdf/wdf-arch.mspx.
[147] David Mosberger and Tai Jin. httperf: A tool for measuring web server performance. In FirstWorkshop on Internet Server Performance, pages 59—67, Madison, WI, June 1998. ACM.
[148] Bob Muglia. Microsoft windows server platform: The next three years. Microsoft Profess-sional Developers Conference Keynote Address, September 2005. Available at http://www.microsoft.com/presspass/exec/bobmuglia/09-15PDC2005.mspx.
[149] Gilles Muller, Michel Banatre, Nadine Peyrouze, and Bruno Rochat. Lessons from FTM: Anexperiment in design and implementation of a low-cost fault-tolerant system. IEEE Transac-tions on Software Engineering, 45(2):332–339, June 1996.
[150] Brendan Murphy. Fault tolerance in this high availability world. Talk given at StanfordUniversity and University of California at Berkeley. Available at http://research.microsoft.com/users/bmurphy/FaultTolerance.htm, October 2000.
[151] Brendan Murphy and Mario R. Garzia. Software reliability engineering for mass marketproducts. The DoD Software Tech News, 8(1), December 2004. Available at http://www.softwaretechnews.com/stn8-1/sremm.html.
[152] Madanlal Musuvathi and Dawson R. Engler. Model checking large network protocol imple-mentations. In Proceedings of the 2004 USENIX Symposium on Network Systems Design andImplementation, pages 155–168, San Francisco, California, March 2004.
150
[153] Madanlal Musuvathi, David Y. W. Park, Andy Chou, Dawson R. Engler, and David L. Dill.CMC: A pragmatic approach to model checking real code. In Proceedings of the 5th USENIXSymposium on Operating Systems Design and Implementation, pages 75–88, Boston, Mas-sachusetts, December 2002.
[154] Nachiappan Nagappan and Thomas Ball. Static analysis tools as early indicators of pre-release defect density. In Proceedings of the 27th International Conference on SoftwareEngineering, pages 580–586, St. Louis, Missouri, May 2005.
[155] Nar Ganapathy, Architect, Windows Device Experience. personal communication, April2005.
[156] Chandra Narayanaswami, Noboru Kamijoh, Mandayam Raghunath, Tadanobu Inoue,Thomas Cipolla, Jim Sanford, Eugene Schlig, Sreekrishnan Venkiteswaran, and Dinakar Gu-niguntala. Ibm’s linux watch: The challenge of miniaturization. IEEE Computer, 35(1),January 2002.
[157] Wee Teck Ng and Peter M. Chen. The systematic improvement of fault tolerance in the Riofile cache. In Proceedings of the 29th Symposium on Fault-Tolerant Computing, pages 76–83,Madison, Wisconsin, June 1999. IEEE.
[158] Walter Oney. Programming the Microsoft Windows Driver Model, Second Edition. MicrosoftPress, Redmond, Washington, December 2002.
[159] Elliott I. Organick. A Programmer’s View of the Intel 432 System. McGraw Hill, 1983.
[160] Vince Orgovan and Mike Tricker. An introduction to driver quality. Microsoft WinHec 2004Presentation DDT301, August 2003.
[161] Alessandro Orso, Anup Rao, and Marj Jean Harrold. A technique for dynamic updating forJava software. In Proceedings of the IEEE International Conference on Software Mainte-nance 2002, pages 649–658, Montreal, Canada, October 2002.
[162] David L. Parnas. On the criteria to be used in decomposing systems into modules. Commu-nications of the ACM, 15(12):1053–1058, December 1972.
[163] David L. Parnas. The influence of software structure on system reliability. In Proceedings ofthe International Converence on Software Engineering, pages 358–362, April 1975.
[164] David Patterson, Aaron Brown, Pete Broadwell, George Candea, Mike Chen, James Cut-ler, Patricia Enriquez, Armando Fox, Emre Kycyman, Matthew Merzbacher, David Oppen-heimer, Naveen Sastry, William Tetzlaff, Jonathan Traupman, and Noah Treuhaft. Recovery-Oriented Computing (ROC): Motivation, definition, techniques, and case studies. TechnicalReport CSD-02-1175, UC Berkeley Computer Science, March 2002.
151
[165] Marc C. Paulk, Bill Curtis, Mary Beth Chrissis, and Charles V. Weber. The capability matu-rity model, version 1.1. IEEE Software, 10(4):18–27, July 1993.
[166] David Pescovitz. Monsters in a box. Wired, 8(12):341–347, December 2000.
[167] James S. Plank, Micah Beck, Gerry Kingsley, and Kai Li. Libckpt: Transparent checkpoint-ing under Unix. In Proceedings of the 1995 Winter USENIX Conference, pages 213–224,New Orleans, Louisiana, January 1995.
[168] Jon Postel. Transmission control protocol. RFC 793, Internet Engineering Task Force,September 1981.
[169] Project-UDI. Introduction to UDI version 1.0. Technical report, Project UDI, August 1999.
[170] Brian Randell. System structure for software fault tolerance. IEEE Transactions on SoftwareEngineering, 1:220–232, June 1975.
[171] Rodrigo Rodrigues, Miguel Castro, and Barbara Liskov. BASE: Using abstraction to improvefault tolerance. In Proceedings of the 18th ACM Symposium on Operating Systems Principles,pages 15–28, Lake Louise, Alberta, October 2001.
[172] Mark Russinovich, Zary Segall, and Dan Siewiorek. Application transparent fault manage-ment in Fault Tolerant Mach. In Proceedings of the 23rd Symposium on Fault-Tolerant Com-puting, pages 10–19, Toulouse, France, June 1993. IEEE.
[173] Mark E. Russionvich and David A. Solomon. Inside Windows 2000. Microsoft Press, Red-mond, Washington, 2000.
[174] Jerome H. Saltzer. Protection and the control of information sharing in Multics. Communi-cations of the ACM, 17(7):388–402, July 1974.
[175] Richard D. Schlichting and Fred B. Schneider. Fail-stop processors: An approach to design-ing fault-tolerant computing systems. ACM Transactions on Computer Systems, 1(3):222–238, August 1983.
[176] Frank Schmuck and Jim Wylie. Experience with transactions in QuickSilver. In Proceedingsof the 13th ACM Symposium on Operating Systems Principles, pages 239–253, Pacific Grove,California, October 1991.
[177] Fred B. Schneider. Byzantine generals in action: Implementing fail-stop processors. ACMTransactions on Computer Systems, 2(2):145–154, May 1984.
[178] Michael D. Schroeder and Michael Burrows. Performance of Firefly RPC. In Proceedingsof the 12th ACM Symposium on Operating Systems Principles, pages 83–90, Litchfield Park,AZ, December 1989.
152
[179] Ephriam Schwartz. Check your oil and your e-mail. PC World, September 1999.
[180] Ephriam Schwartz. Linux emerges as cell phone dark horse. Inforworld, October 2003. Avail-able at http://www.infoworld.com/article/03/10/03/39NNphoneos 1.html.
[181] Mark E. Segal and Ophir Frieder. On-the-fly program modification: Systems for dynamicupdating. IEEE Software, 2(10):53–65, March 1993.
[182] Margo I. Seltzer, Yasuhiro Endo, Christopher Small, and Keith A. Smith. Dealing with disas-ter: Surviving misbehaved kernel extensions. In Proceedings of the 2nd USENIX Symposiumon Operating Systems Design and Implementation, pages 213–227, Seattle, Washington, Oc-tober 1996.
[183] Mehul A. Shah, Joseph M. Hellerstein, and Eric Brewer. Highly available, fault-tolerant,parallel dataflows. In Proceedings of the 2004 ACM SIGMOD International Conference onManagement of Data, pages 827–838, Paris, France, June 2004.
[184] Sharp. Zaurus handheld SL-6000, 2005. http://www.sharpusa.com/products/TypeLanding/0,1056,112,00.html.
[185] Takahiro Shinagawa, Kenji Kono, and Takashi Masuda. Exploiting segmentation mechanismfor protecting against malicious mobile code. Technical Report 00-02, Dept. of InformationScience, University of Tokyo, May 2000.
[186] Abraham Silberschatz, Peter Baer Galvin, and Greg Gagne. Operating System Concepts,Seventh Edition. John Wiley & Sons, Inc., March 2004.
[187] Slashdot. Windows upgrade, FAA error cause LAX shutdown. http://it.slashdot.org/it/04/09/21/2120203.shtml?tid=128&tid=103&tid=201.
[188] Craig A. N. Soules, Jonathan Appavoo, Kevin Hui, Robert W. Wisniewski, Dilma Da Silva,Gregory R. Ganger, Orran Krieger, Michael Stumm, Marc Auslander, Michal Ostrowski,Bryan Rosenburg, and Jimi Xenidis. System support for online reconfiguration. In Proceed-ings of the 2003 USENIX Annual Technical Conference, pages 141–154, Austin, Texas, June2003.
[189] Standard Performance Evaluation Corporation. The SPECweb99 benchmark, 1999.
[190] Jeremy Sugerman, Ganesh Venkitachalam, and Beng-Hong Lim. Virtualizing I/O devices onVMware workstation’s hosted virtual machine monitor. In Proceedings of the 2001 USENIXAnnual Technical Conference, pages 1–14, Boston, Massachusetts, June 2001.
153
[191] Mark Sullivan and Ram Chillarege. Software defects and their impact on system availability- a study of field failures in operating systems. In Proceedings of the 21st Symposium onFault-Tolerant Computing, pages 2–9, Quebec, Canada, June 1991. IEEE.
[192] Mark Sullivan and Michael Stonebraker. Using write protected data structures to improvesoftware fault tolerance in highly available database management systems. In Proceedingsof the 17th International Conference on Very Large Data Bases, pages 171–180, Barcelona,Spain, September 1991. Morgan Kaufman Publishing.
[193] Paul Thurrott. Windows 2000 server: The road to gold, part two: Developing Windows. PaulThurrott’s SuperSite for Windows, January 2003.
[194] TiVo Corporation. TiVo digital video recorder, 2001. www.tivo.com.
[195] Wilfredo Torres-Pomales. Software fault tolerance: A tutorial. Technical Report Tm-2000-210616, NASA Langley Research Center, October 2000.
[196] Linux Torvalds. Linux kernel source tree. Available at http://www.kernel.org.
[197] Jeff Tranter. Open sound system programmer’s guide, January 2000. Available at http://www.opensound.com/pguide/oss.pdf.
[198] Harvey Tuch, Gerwin Klein, and Gernot Heiser. OS verification – now! In Proceedings ofthe Tenth IEEE Workshop on Hot Topics in Operating Systems, Santa Fe, New Mexico, June2005.
[199] Alan Turing. On computable numbers, with an application to the entscheidungsproblem.London Math Society, Series 2, 42(2):230–265, November 1936.
[200] Arjan van de Ven. kHTTPd: Linux HTTP accelerator. Available at http://www.fenrus.demon.nl/.
[201] Kevin Thomas Van Maren. The Fluke device driver framework. Master’s thesis, Departmentof Computer Science, University of Utah, December 1999.
[202] Werner Vogels, Dan Dumitriu, Ken Birman Rod Gamache, Mike Massa, Rob Short, JohnVert, Joe Barrera, and Jim Gray. The design and architecture of the microsoft cluster ser-vice. In Proceedings of the 28th Symposium on Fault-Tolerant Computing, pages 422–431,Munich, Germany, June 1998. IEEE.
[203] Thorsten von Eicken, David E. Culler, Seth Copen Goldstein, and Klaus Erik Schauser. Ac-tive messages: a mechanism for integrated communication and computation. In Proceed-ings of the 19th ACM International Symposium on Computer Architecture, pages 256–266,Queensland, Australia, May 1992.
154
[204] John von Neumann. Probabilistic logics and synthesis of reliable organisms from unreliablecomponents. In Claude Shannon and John McCarthy, editors, Automata Studies, pages 43–98. Princeton University Press, Princeton, New Jersey, 1956.
[205] Robert S. Wahbe, Steven E. Lucco, Thomas E. Anderson, and Susan L. Graham. Efficientsoftware-based fault isolation. In Proceedings of the 14th ACM Symposium on OperatingSystems Principles, pages 203–216, Asheville, North Carolina, December 1993.
[206] David A. Wheeler. More than a gigabuck: Estimating GNU/Linux’s size,July 2002. Available at http://www.dwheeler.com/sloc/redhat71-v1/redhat71sloc.html.
[207] Andrew Whitaker, Marianne Shaw, and Steven D. Gribble. Denali: Lightweight virtual ma-chines for distributed and networked applications. In Proceedings of the 5th USENIX Sym-posium on Operating Systems Design and Implementation, pages 195–209, Boston, Mas-sachusetts, December 2002.
[208] Toh Ne Win and Michael Ernst. Verifying distributed algorithms via dynamic analysis andtheorem proving. Technical Report MIT-LCS-TR-841, MIT Lab for Computer Science, May2002.
[209] Emmett Witchel, Josh Cates, and Krste Asanovic. Mondrian memory protection. In Proceed-ings of the Tenth ACM International Conference on Architectural Support for ProgrammingLanguages and Operating Systems, pages 304–316, October 2002.
[210] William A. Wulf. Reliable hardware-software architecture. IEEE Transactions on SoftwareEngineering, 1(2):233–240, June 1975.
[211] Junfeng Yang, Paul Twohey, Dawson Engler, and Madanlal Musuvathi. Using model check-ing to find serious file systems errors. In Proceedings of the 6th USENIX Symposium onOperating Systems Design and Implementation, pages 273–288, San Francisco, California,December 2004.
[212] Michael Young, Mike Accetta, Robert Baron, William Bolosky, David Golub, RichardRashid, and Avadis Tevanian. Mach: A new kernel foundation for UNIX development. InProceedings of the 1986 Summer USENIX Conference, pages 93–113, Atlanta, Georgia, June1986.
155
VITA
Michael Swift was born and raised in Amherst, Massachusetts. After graduating from Cornell
University in 1992, he moved to the Northwest and started working for Microsoft. There he im-
plemented security features in Windows NT 4.0 and Windows 2000, primarily access control and
authentication code. However, his lifelong dream was to become a computer science professor. The
expiration of his G.R.E. scores in 1996 provided the final kick in the pants that got him out of Mi-
crosoft. Michael applied to graduate school that year and started at the University of Washington
two years later. He earned a Doctor of Philosophy in Computer Science from the University of
Washington in 2005.